Initial commit

comfy/ldm/modules/distributions/distributions.py

+import torch
+import numpy as np
+
+
+class AbstractDistribution:
+    def sample(self):
+        raise NotImplementedError()
+
+    def mode(self):
+        raise NotImplementedError()
+
+
+class DiracDistribution(AbstractDistribution):
+    def __init__(self, value):
+        self.value = value
+
+    def sample(self):
+        return self.value
+
+    def mode(self):
+        return self.value
+
+
+class DiagonalGaussianDistribution(object):
+    def __init__(self, parameters, deterministic=False):
+        self.parameters = parameters
+        self.mean, self.logvar = torch.chunk(parameters, 2, dim=1)
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.deterministic = deterministic
+        self.std = torch.exp(0.5 * self.logvar)
+        self.var = torch.exp(self.logvar)
+        if self.deterministic:
+            self.var = self.std = torch.zeros_like(self.mean, device=self.parameters.device)
+
+    def sample(self):
+        x = self.mean + self.std * torch.randn(self.mean.shape, device=self.parameters.device)
+        return x
+
+    def kl(self, other=None):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        else:
+            if other is None:
+                return 0.5 * torch.sum(torch.pow(self.mean, 2)
+                                       + self.var - 1.0 - self.logvar,
+                                       dim=[1, 2, 3])
+            else:
+                return 0.5 * torch.sum(
+                    torch.pow(self.mean - other.mean, 2) / other.var
+                    + self.var / other.var - 1.0 - self.logvar + other.logvar,
+                    dim=[1, 2, 3])
+
+    def nll(self, sample, dims=[1,2,3]):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * torch.sum(
+            logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
+            dim=dims)
+
+    def mode(self):
+        return self.mean
+
+
+def normal_kl(mean1, logvar1, mean2, logvar2):
+    """
+    source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
+    Compute the KL divergence between two gaussians.
+    Shapes are automatically broadcasted, so batches can be compared to
+    scalars, among other use cases.
+    """
+    tensor = None
+    for obj in (mean1, logvar1, mean2, logvar2):
+        if isinstance(obj, torch.Tensor):
+            tensor = obj
+            break
+    assert tensor is not None, "at least one argument must be a Tensor"
+
+    # Force variances to be Tensors. Broadcasting helps convert scalars to
+    # Tensors, but it does not work for torch.exp().
+    logvar1, logvar2 = [
+        x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
+        for x in (logvar1, logvar2)
+    ]
+
+    return 0.5 * (
+        -1.0
+        + logvar2
+        - logvar1
+        + torch.exp(logvar1 - logvar2)
+        + ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
+    )

comfy/ldm/modules/ema.py

+import torch
+from torch import nn
+
+
+class LitEma(nn.Module):
+    def __init__(self, model, decay=0.9999, use_num_upates=True):
+        super().__init__()
+        if decay < 0.0 or decay > 1.0:
+            raise ValueError('Decay must be between 0 and 1')
+
+        self.m_name2s_name = {}
+        self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32))
+        self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int) if use_num_upates
+        else torch.tensor(-1, dtype=torch.int))
+
+        for name, p in model.named_parameters():
+            if p.requires_grad:
+                # remove as '.'-character is not allowed in buffers
+                s_name = name.replace('.', '')
+                self.m_name2s_name.update({name: s_name})
+                self.register_buffer(s_name, p.clone().detach().data)
+
+        self.collected_params = []
+
+    def reset_num_updates(self):
+        del self.num_updates
+        self.register_buffer('num_updates', torch.tensor(0, dtype=torch.int))
+
+    def forward(self, model):
+        decay = self.decay
+
+        if self.num_updates >= 0:
+            self.num_updates += 1
+            decay = min(self.decay, (1 + self.num_updates) / (10 + self.num_updates))
+
+        one_minus_decay = 1.0 - decay
+
+        with torch.no_grad():
+            m_param = dict(model.named_parameters())
+            shadow_params = dict(self.named_buffers())
+
+            for key in m_param:
+                if m_param[key].requires_grad:
+                    sname = self.m_name2s_name[key]
+                    shadow_params[sname] = shadow_params[sname].type_as(m_param[key])
+                    shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key]))
+                else:
+                    assert not key in self.m_name2s_name
+
+    def copy_to(self, model):
+        m_param = dict(model.named_parameters())
+        shadow_params = dict(self.named_buffers())
+        for key in m_param:
+            if m_param[key].requires_grad:
+                m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data)
+            else:
+                assert not key in self.m_name2s_name
+
+    def store(self, parameters):
+        """
+        Save the current parameters for restoring later.
+        Args:
+          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+            temporarily stored.
+        """
+        self.collected_params = [param.clone() for param in parameters]
+
+    def restore(self, parameters):
+        """
+        Restore the parameters stored with the `store` method.
+        Useful to validate the model with EMA parameters without affecting the
+        original optimization process. Store the parameters before the
+        `copy_to` method. After validation (or model saving), use this to
+        restore the former parameters.
+        Args:
+          parameters: Iterable of `torch.nn.Parameter`; the parameters to be
+            updated with the stored parameters.
+        """
+        for c_param, param in zip(self.collected_params, parameters):
+            param.data.copy_(c_param.data)

comfy/ldm/modules/encoders/noise_aug_modules.py

+from ..diffusionmodules.upscaling import ImageConcatWithNoiseAugmentation
+from ..diffusionmodules.openaimodel import Timestep
+import torch
+
+class CLIPEmbeddingNoiseAugmentation(ImageConcatWithNoiseAugmentation):
+    def __init__(self, *args, clip_stats_path=None, timestep_dim=256, **kwargs):
+        super().__init__(*args, **kwargs)
+        if clip_stats_path is None:
+            clip_mean, clip_std = torch.zeros(timestep_dim), torch.ones(timestep_dim)
+        else:
+            clip_mean, clip_std = torch.load(clip_stats_path, map_location="cpu")
+        self.register_buffer("data_mean", clip_mean[None, :], persistent=False)
+        self.register_buffer("data_std", clip_std[None, :], persistent=False)
+        self.time_embed = Timestep(timestep_dim)
+
+    def scale(self, x):
+        # re-normalize to centered mean and unit variance
+        x = (x - self.data_mean.to(x.device)) * 1. / self.data_std.to(x.device)
+        return x
+
+    def unscale(self, x):
+        # back to original data stats
+        x = (x * self.data_std.to(x.device)) + self.data_mean.to(x.device)
+        return x
+
+    def forward(self, x, noise_level=None, seed=None):
+        if noise_level is None:
+            noise_level = torch.randint(0, self.max_noise_level, (x.shape[0],), device=x.device).long()
+        else:
+            assert isinstance(noise_level, torch.Tensor)
+        x = self.scale(x)
+        z = self.q_sample(x, noise_level, seed=seed)
+        z = self.unscale(z)
+        noise_level = self.time_embed(noise_level)
+        return z, noise_level

comfy/ldm/modules/sub_quadratic_attention.py

+# original source:
+#   https://github.com/AminRezaei0x443/memory-efficient-attention/blob/1bc0d9e6ac5f82ea43a375135c4e1d3896ee1694/memory_efficient_attention/attention_torch.py
+# license:
+#   MIT
+# credit:
+#   Amin Rezaei (original author)
+#   Alex Birch (optimized algorithm for 3D tensors, at the expense of removing bias, masking and callbacks)
+# implementation of:
+#   Self-attention Does Not Need O(n2) Memory":
+#   https://arxiv.org/abs/2112.05682v2
+
+from functools import partial
+import torch
+from torch import Tensor
+from torch.utils.checkpoint import checkpoint
+import math
+import logging
+
+try:
+    from typing import Optional, NamedTuple, List, Protocol
+except ImportError:
+    from typing import Optional, NamedTuple, List
+    from typing_extensions import Protocol
+
+from typing import List
+
+from comfy import model_management
+
+def dynamic_slice(
+    x: Tensor,
+    starts: List[int],
+    sizes: List[int],
+) -> Tensor:
+    slicing = tuple(slice(start, start + size) for start, size in zip(starts, sizes))
+    return x[slicing]
+
+class AttnChunk(NamedTuple):
+    exp_values: Tensor
+    exp_weights_sum: Tensor
+    max_score: Tensor
+
+class SummarizeChunk(Protocol):
+    @staticmethod
+    def __call__(
+        query: Tensor,
+        key_t: Tensor,
+        value: Tensor,
+    ) -> AttnChunk: ...
+
+class ComputeQueryChunkAttn(Protocol):
+    @staticmethod
+    def __call__(
+        query: Tensor,
+        key_t: Tensor,
+        value: Tensor,
+    ) -> Tensor: ...
+
+def _summarize_chunk(
+    query: Tensor,
+    key_t: Tensor,
+    value: Tensor,
+    scale: float,
+    upcast_attention: bool,
+    mask,
+) -> AttnChunk:
+    if upcast_attention:
+        with torch.autocast(enabled=False, device_type = 'cuda'):
+            query = query.float()
+            key_t = key_t.float()
+            attn_weights = torch.baddbmm(
+                torch.empty(1, 1, 1, device=query.device, dtype=query.dtype),
+                query,
+                key_t,
+                alpha=scale,
+                beta=0,
+            )
+    else:
+        attn_weights = torch.baddbmm(
+            torch.empty(1, 1, 1, device=query.device, dtype=query.dtype),
+            query,
+            key_t,
+            alpha=scale,
+            beta=0,
+        )
+    max_score, _ = torch.max(attn_weights, -1, keepdim=True)
+    max_score = max_score.detach()
+    attn_weights -= max_score
+    if mask is not None:
+        attn_weights += mask
+    torch.exp(attn_weights, out=attn_weights)
+    exp_weights = attn_weights.to(value.dtype)
+    exp_values = torch.bmm(exp_weights, value)
+    max_score = max_score.squeeze(-1)
+    return AttnChunk(exp_values, exp_weights.sum(dim=-1), max_score)
+
+def _query_chunk_attention(
+    query: Tensor,
+    key_t: Tensor,
+    value: Tensor,
+    summarize_chunk: SummarizeChunk,
+    kv_chunk_size: int,
+    mask,
+) -> Tensor:
+    batch_x_heads, k_channels_per_head, k_tokens = key_t.shape
+    _, _, v_channels_per_head = value.shape
+
+    def chunk_scanner(chunk_idx: int, mask) -> AttnChunk:
+        key_chunk = dynamic_slice(
+            key_t,
+            (0, 0, chunk_idx),
+            (batch_x_heads, k_channels_per_head, kv_chunk_size)
+        )
+        value_chunk = dynamic_slice(
+            value,
+            (0, chunk_idx, 0),
+            (batch_x_heads, kv_chunk_size, v_channels_per_head)
+        )
+        if mask is not None:
+            mask = mask[:,:,chunk_idx:chunk_idx + kv_chunk_size]
+
+        return summarize_chunk(query, key_chunk, value_chunk, mask=mask)
+
+    chunks: List[AttnChunk] = [
+        chunk_scanner(chunk, mask) for chunk in torch.arange(0, k_tokens, kv_chunk_size)
+    ]
+    acc_chunk = AttnChunk(*map(torch.stack, zip(*chunks)))
+    chunk_values, chunk_weights, chunk_max = acc_chunk
+
+    global_max, _ = torch.max(chunk_max, 0, keepdim=True)
+    max_diffs = torch.exp(chunk_max - global_max)
+    chunk_values *= torch.unsqueeze(max_diffs, -1)
+    chunk_weights *= max_diffs
+
+    all_values = chunk_values.sum(dim=0)
+    all_weights = torch.unsqueeze(chunk_weights, -1).sum(dim=0)
+    return all_values / all_weights
+
+# TODO: refactor CrossAttention#get_attention_scores to share code with this
+def _get_attention_scores_no_kv_chunking(
+    query: Tensor,
+    key_t: Tensor,
+    value: Tensor,
+    scale: float,
+    upcast_attention: bool,
+    mask,
+) -> Tensor:
+    if upcast_attention:
+        with torch.autocast(enabled=False, device_type = 'cuda'):
+            query = query.float()
+            key_t = key_t.float()
+            attn_scores = torch.baddbmm(
+                torch.empty(1, 1, 1, device=query.device, dtype=query.dtype),
+                query,
+                key_t,
+                alpha=scale,
+                beta=0,
+            )
+    else:
+        attn_scores = torch.baddbmm(
+            torch.empty(1, 1, 1, device=query.device, dtype=query.dtype),
+            query,
+            key_t,
+            alpha=scale,
+            beta=0,
+        )
+
+    if mask is not None:
+        attn_scores += mask
+    try:
+        attn_probs = attn_scores.softmax(dim=-1)
+        del attn_scores
+    except model_management.OOM_EXCEPTION:
+        logging.warning("ran out of memory while running softmax in  _get_attention_scores_no_kv_chunking, trying slower in place softmax instead")
+        attn_scores -= attn_scores.max(dim=-1, keepdim=True).values # noqa: F821 attn_scores is not defined
+        torch.exp(attn_scores, out=attn_scores)
+        summed = torch.sum(attn_scores, dim=-1, keepdim=True)
+        attn_scores /= summed
+        attn_probs = attn_scores
+
+    hidden_states_slice = torch.bmm(attn_probs.to(value.dtype), value)
+    return hidden_states_slice
+
+class ScannedChunk(NamedTuple):
+    chunk_idx: int
+    attn_chunk: AttnChunk
+
+def efficient_dot_product_attention(
+    query: Tensor,
+    key_t: Tensor,
+    value: Tensor,
+    query_chunk_size=1024,
+    kv_chunk_size: Optional[int] = None,
+    kv_chunk_size_min: Optional[int] = None,
+    use_checkpoint=True,
+    upcast_attention=False,
+    mask = None,
+):
+    """Computes efficient dot-product attention given query, transposed key, and value.
+      This is efficient version of attention presented in
+      https://arxiv.org/abs/2112.05682v2 which comes with O(sqrt(n)) memory requirements.
+      Args:
+        query: queries for calculating attention with shape of
+          `[batch * num_heads, tokens, channels_per_head]`.
+        key_t: keys for calculating attention with shape of
+          `[batch * num_heads, channels_per_head, tokens]`.
+        value: values to be used in attention with shape of
+          `[batch * num_heads, tokens, channels_per_head]`.
+        query_chunk_size: int: query chunks size
+        kv_chunk_size: Optional[int]: key/value chunks size. if None: defaults to sqrt(key_tokens)
+        kv_chunk_size_min: Optional[int]: key/value minimum chunk size. only considered when kv_chunk_size is None. changes `sqrt(key_tokens)` into `max(sqrt(key_tokens), kv_chunk_size_min)`, to ensure our chunk sizes don't get too small (smaller chunks = more chunks = less concurrent work done).
+        use_checkpoint: bool: whether to use checkpointing (recommended True for training, False for inference)
+      Returns:
+        Output of shape `[batch * num_heads, query_tokens, channels_per_head]`.
+      """
+    batch_x_heads, q_tokens, q_channels_per_head = query.shape
+    _, _, k_tokens = key_t.shape
+    scale = q_channels_per_head ** -0.5
+
+    kv_chunk_size = min(kv_chunk_size or int(math.sqrt(k_tokens)), k_tokens)
+    if kv_chunk_size_min is not None:
+        kv_chunk_size = max(kv_chunk_size, kv_chunk_size_min)
+
+    if mask is not None and len(mask.shape) == 2:
+        mask = mask.unsqueeze(0)
+
+    def get_query_chunk(chunk_idx: int) -> Tensor:
+        return dynamic_slice(
+            query,
+            (0, chunk_idx, 0),
+            (batch_x_heads, min(query_chunk_size, q_tokens), q_channels_per_head)
+        )
+
+    def get_mask_chunk(chunk_idx: int) -> Tensor:
+        if mask is None:
+            return None
+        if mask.shape[1] == 1:
+            return mask
+        chunk = min(query_chunk_size, q_tokens)
+        return mask[:,chunk_idx:chunk_idx + chunk]
+
+    summarize_chunk: SummarizeChunk = partial(_summarize_chunk, scale=scale, upcast_attention=upcast_attention)
+    summarize_chunk: SummarizeChunk = partial(checkpoint, summarize_chunk) if use_checkpoint else summarize_chunk
+    compute_query_chunk_attn: ComputeQueryChunkAttn = partial(
+        _get_attention_scores_no_kv_chunking,
+        scale=scale,
+        upcast_attention=upcast_attention
+    ) if k_tokens <= kv_chunk_size else (
+        # fast-path for when there's just 1 key-value chunk per query chunk (this is just sliced attention btw)
+        partial(
+            _query_chunk_attention,
+            kv_chunk_size=kv_chunk_size,
+            summarize_chunk=summarize_chunk,
+        )
+    )
+
+    if q_tokens <= query_chunk_size:
+        # fast-path for when there's just 1 query chunk
+        return compute_query_chunk_attn(
+            query=query,
+            key_t=key_t,
+            value=value,
+            mask=mask,
+        )
+
+    # TODO: maybe we should use torch.empty_like(query) to allocate storage in-advance,
+    # and pass slices to be mutated, instead of torch.cat()ing the returned slices
+    res = torch.cat([
+        compute_query_chunk_attn(
+            query=get_query_chunk(i * query_chunk_size),
+            key_t=key_t,
+            value=value,
+            mask=get_mask_chunk(i * query_chunk_size)
+        ) for i in range(math.ceil(q_tokens / query_chunk_size))
+    ], dim=1)
+    return res

comfy/ldm/modules/temporal_ae.py

+import functools
+from typing import Iterable, Union
+
+import torch
+from einops import rearrange, repeat
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+from .diffusionmodules.model import (
+    AttnBlock,
+    Decoder,
+    ResnetBlock,
+)
+from .diffusionmodules.openaimodel import ResBlock, timestep_embedding
+from .attention import BasicTransformerBlock
+
+def partialclass(cls, *args, **kwargs):
+    class NewCls(cls):
+        __init__ = functools.partialmethod(cls.__init__, *args, **kwargs)
+
+    return NewCls
+
+
+class VideoResBlock(ResnetBlock):
+    def __init__(
+        self,
+        out_channels,
+        *args,
+        dropout=0.0,
+        video_kernel_size=3,
+        alpha=0.0,
+        merge_strategy="learned",
+        **kwargs,
+    ):
+        super().__init__(out_channels=out_channels, dropout=dropout, *args, **kwargs)
+        if video_kernel_size is None:
+            video_kernel_size = [3, 1, 1]
+        self.time_stack = ResBlock(
+            channels=out_channels,
+            emb_channels=0,
+            dropout=dropout,
+            dims=3,
+            use_scale_shift_norm=False,
+            use_conv=False,
+            up=False,
+            down=False,
+            kernel_size=video_kernel_size,
+            use_checkpoint=False,
+            skip_t_emb=True,
+        )
+
+        self.merge_strategy = merge_strategy
+        if self.merge_strategy == "fixed":
+            self.register_buffer("mix_factor", torch.Tensor([alpha]))
+        elif self.merge_strategy == "learned":
+            self.register_parameter(
+                "mix_factor", torch.nn.Parameter(torch.Tensor([alpha]))
+            )
+        else:
+            raise ValueError(f"unknown merge strategy {self.merge_strategy}")
+
+    def get_alpha(self, bs):
+        if self.merge_strategy == "fixed":
+            return self.mix_factor
+        elif self.merge_strategy == "learned":
+            return torch.sigmoid(self.mix_factor)
+        else:
+            raise NotImplementedError()
+
+    def forward(self, x, temb, skip_video=False, timesteps=None):
+        b, c, h, w = x.shape
+        if timesteps is None:
+            timesteps = b
+
+        x = super().forward(x, temb)
+
+        if not skip_video:
+            x_mix = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps)
+
+            x = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps)
+
+            x = self.time_stack(x, temb)
+
+            alpha = self.get_alpha(bs=b // timesteps).to(x.device)
+            x = alpha * x + (1.0 - alpha) * x_mix
+
+            x = rearrange(x, "b c t h w -> (b t) c h w")
+        return x
+
+
+class AE3DConv(ops.Conv2d):
+    def __init__(self, in_channels, out_channels, video_kernel_size=3, *args, **kwargs):
+        super().__init__(in_channels, out_channels, *args, **kwargs)
+        if isinstance(video_kernel_size, Iterable):
+            padding = [int(k // 2) for k in video_kernel_size]
+        else:
+            padding = int(video_kernel_size // 2)
+
+        self.time_mix_conv = ops.Conv3d(
+            in_channels=out_channels,
+            out_channels=out_channels,
+            kernel_size=video_kernel_size,
+            padding=padding,
+        )
+
+    def forward(self, input, timesteps=None, skip_video=False):
+        if timesteps is None:
+            timesteps = input.shape[0]
+        x = super().forward(input)
+        if skip_video:
+            return x
+        x = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps)
+        x = self.time_mix_conv(x)
+        return rearrange(x, "b c t h w -> (b t) c h w")
+
+
+class AttnVideoBlock(AttnBlock):
+    def __init__(
+        self, in_channels: int, alpha: float = 0, merge_strategy: str = "learned"
+    ):
+        super().__init__(in_channels)
+        # no context, single headed, as in base class
+        self.time_mix_block = BasicTransformerBlock(
+            dim=in_channels,
+            n_heads=1,
+            d_head=in_channels,
+            checkpoint=False,
+            ff_in=True,
+        )
+
+        time_embed_dim = self.in_channels * 4
+        self.video_time_embed = torch.nn.Sequential(
+            ops.Linear(self.in_channels, time_embed_dim),
+            torch.nn.SiLU(),
+            ops.Linear(time_embed_dim, self.in_channels),
+        )
+
+        self.merge_strategy = merge_strategy
+        if self.merge_strategy == "fixed":
+            self.register_buffer("mix_factor", torch.Tensor([alpha]))
+        elif self.merge_strategy == "learned":
+            self.register_parameter(
+                "mix_factor", torch.nn.Parameter(torch.Tensor([alpha]))
+            )
+        else:
+            raise ValueError(f"unknown merge strategy {self.merge_strategy}")
+
+    def forward(self, x, timesteps=None, skip_time_block=False):
+        if skip_time_block:
+            return super().forward(x)
+
+        if timesteps is None:
+            timesteps = x.shape[0]
+
+        x_in = x
+        x = self.attention(x)
+        h, w = x.shape[2:]
+        x = rearrange(x, "b c h w -> b (h w) c")
+
+        x_mix = x
+        num_frames = torch.arange(timesteps, device=x.device)
+        num_frames = repeat(num_frames, "t -> b t", b=x.shape[0] // timesteps)
+        num_frames = rearrange(num_frames, "b t -> (b t)")
+        t_emb = timestep_embedding(num_frames, self.in_channels, repeat_only=False)
+        emb = self.video_time_embed(t_emb)  # b, n_channels
+        emb = emb[:, None, :]
+        x_mix = x_mix + emb
+
+        alpha = self.get_alpha().to(x.device)
+        x_mix = self.time_mix_block(x_mix, timesteps=timesteps)
+        x = alpha * x + (1.0 - alpha) * x_mix  # alpha merge
+
+        x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w)
+        x = self.proj_out(x)
+
+        return x_in + x
+
+    def get_alpha(
+        self,
+    ):
+        if self.merge_strategy == "fixed":
+            return self.mix_factor
+        elif self.merge_strategy == "learned":
+            return torch.sigmoid(self.mix_factor)
+        else:
+            raise NotImplementedError(f"unknown merge strategy {self.merge_strategy}")
+
+
+
+def make_time_attn(
+    in_channels,
+    attn_type="vanilla",
+    attn_kwargs=None,
+    alpha: float = 0,
+    merge_strategy: str = "learned",
+    conv_op=ops.Conv2d,
+):
+    return partialclass(
+        AttnVideoBlock, in_channels, alpha=alpha, merge_strategy=merge_strategy
+    )
+
+
+class Conv2DWrapper(torch.nn.Conv2d):
+    def forward(self, input: torch.Tensor, **kwargs) -> torch.Tensor:
+        return super().forward(input)
+
+
+class VideoDecoder(Decoder):
+    available_time_modes = ["all", "conv-only", "attn-only"]
+
+    def __init__(
+        self,
+        *args,
+        video_kernel_size: Union[int, list] = 3,
+        alpha: float = 0.0,
+        merge_strategy: str = "learned",
+        time_mode: str = "conv-only",
+        **kwargs,
+    ):
+        self.video_kernel_size = video_kernel_size
+        self.alpha = alpha
+        self.merge_strategy = merge_strategy
+        self.time_mode = time_mode
+        assert (
+            self.time_mode in self.available_time_modes
+        ), f"time_mode parameter has to be in {self.available_time_modes}"
+
+        if self.time_mode != "attn-only":
+            kwargs["conv_out_op"] = partialclass(AE3DConv, video_kernel_size=self.video_kernel_size)
+        if self.time_mode not in ["conv-only", "only-last-conv"]:
+            kwargs["attn_op"] = partialclass(make_time_attn, alpha=self.alpha, merge_strategy=self.merge_strategy)
+        if self.time_mode not in ["attn-only", "only-last-conv"]:
+            kwargs["resnet_op"] = partialclass(VideoResBlock, video_kernel_size=self.video_kernel_size, alpha=self.alpha, merge_strategy=self.merge_strategy)
+
+        super().__init__(*args, **kwargs)
+
+    def get_last_layer(self, skip_time_mix=False, **kwargs):
+        if self.time_mode == "attn-only":
+            raise NotImplementedError("TODO")
+        else:
+            return (
+                self.conv_out.time_mix_conv.weight
+                if not skip_time_mix
+                else self.conv_out.weight
+            )

comfy/ldm/omnigen/omnigen2.py

+# Original code: https://github.com/VectorSpaceLab/OmniGen2
+
+from typing import Optional, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange, repeat
+from comfy.ldm.lightricks.model import Timesteps
+from comfy.ldm.flux.layers import EmbedND
+from comfy.ldm.modules.attention import optimized_attention_masked
+import comfy.model_management
+import comfy.ldm.common_dit
+
+
+def apply_rotary_emb(x, freqs_cis):
+    if x.shape[1] == 0:
+        return x
+
+    t_ = x.reshape(*x.shape[:-1], -1, 1, 2)
+    t_out = freqs_cis[..., 0] * t_[..., 0] + freqs_cis[..., 1] * t_[..., 1]
+    return t_out.reshape(*x.shape).to(dtype=x.dtype)
+
+
+def swiglu(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor:
+    return F.silu(x) * y
+
+
+class TimestepEmbedding(nn.Module):
+    def __init__(self, in_channels: int, time_embed_dim: int, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.linear_1 = operations.Linear(in_channels, time_embed_dim, dtype=dtype, device=device)
+        self.act = nn.SiLU()
+        self.linear_2 = operations.Linear(time_embed_dim, time_embed_dim, dtype=dtype, device=device)
+
+    def forward(self, sample: torch.Tensor) -> torch.Tensor:
+        sample = self.linear_1(sample)
+        sample = self.act(sample)
+        sample = self.linear_2(sample)
+        return sample
+
+
+class LuminaRMSNormZero(nn.Module):
+    def __init__(self, embedding_dim: int, norm_eps: float = 1e-5, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(min(embedding_dim, 1024), 4 * embedding_dim, dtype=dtype, device=device)
+        self.norm = operations.RMSNorm(embedding_dim, eps=norm_eps, dtype=dtype, device=device)
+
+    def forward(self, x: torch.Tensor, emb: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        emb = self.linear(self.silu(emb))
+        scale_msa, gate_msa, scale_mlp, gate_mlp = emb.chunk(4, dim=1)
+        x = self.norm(x) * (1 + scale_msa[:, None])
+        return x, gate_msa, scale_mlp, gate_mlp
+
+
+class LuminaLayerNormContinuous(nn.Module):
+    def __init__(self, embedding_dim: int, conditioning_embedding_dim: int, elementwise_affine: bool = False, eps: float = 1e-6, out_dim: Optional[int] = None, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear_1 = operations.Linear(conditioning_embedding_dim, embedding_dim, dtype=dtype, device=device)
+        self.norm = operations.LayerNorm(embedding_dim, eps, elementwise_affine, dtype=dtype, device=device)
+        self.linear_2 = operations.Linear(embedding_dim, out_dim, bias=True, dtype=dtype, device=device) if out_dim is not None else None
+
+    def forward(self, x: torch.Tensor, conditioning_embedding: torch.Tensor) -> torch.Tensor:
+        emb = self.linear_1(self.silu(conditioning_embedding).to(x.dtype))
+        x = self.norm(x) * (1 + emb)[:, None, :]
+        if self.linear_2 is not None:
+            x = self.linear_2(x)
+        return x
+
+
+class LuminaFeedForward(nn.Module):
+    def __init__(self, dim: int, inner_dim: int, multiple_of: int = 256, dtype=None, device=None, operations=None):
+        super().__init__()
+        inner_dim = multiple_of * ((inner_dim + multiple_of - 1) // multiple_of)
+        self.linear_1 = operations.Linear(dim, inner_dim, bias=False, dtype=dtype, device=device)
+        self.linear_2 = operations.Linear(inner_dim, dim, bias=False, dtype=dtype, device=device)
+        self.linear_3 = operations.Linear(dim, inner_dim, bias=False, dtype=dtype, device=device)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        h1, h2 = self.linear_1(x), self.linear_3(x)
+        return self.linear_2(swiglu(h1, h2))
+
+
+class Lumina2CombinedTimestepCaptionEmbedding(nn.Module):
+    def __init__(self, hidden_size: int = 4096, text_feat_dim: int = 2048, frequency_embedding_size: int = 256, norm_eps: float = 1e-5, timestep_scale: float = 1.0, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.time_proj = Timesteps(num_channels=frequency_embedding_size, flip_sin_to_cos=True, downscale_freq_shift=0.0, scale=timestep_scale)
+        self.timestep_embedder = TimestepEmbedding(in_channels=frequency_embedding_size, time_embed_dim=min(hidden_size, 1024), dtype=dtype, device=device, operations=operations)
+        self.caption_embedder = nn.Sequential(
+            operations.RMSNorm(text_feat_dim, eps=norm_eps, dtype=dtype, device=device),
+            operations.Linear(text_feat_dim, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+
+    def forward(self, timestep: torch.Tensor, text_hidden_states: torch.Tensor, dtype: torch.dtype) -> Tuple[torch.Tensor, torch.Tensor]:
+        timestep_proj = self.time_proj(timestep).to(dtype=dtype)
+        time_embed = self.timestep_embedder(timestep_proj)
+        caption_embed = self.caption_embedder(text_hidden_states)
+        return time_embed, caption_embed
+
+
+class Attention(nn.Module):
+    def __init__(self, query_dim: int, dim_head: int, heads: int, kv_heads: int, eps: float = 1e-5, bias: bool = False, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.heads = heads
+        self.kv_heads = kv_heads
+        self.dim_head = dim_head
+        self.scale = dim_head ** -0.5
+
+        self.to_q = operations.Linear(query_dim, heads * dim_head, bias=bias, dtype=dtype, device=device)
+        self.to_k = operations.Linear(query_dim, kv_heads * dim_head, bias=bias, dtype=dtype, device=device)
+        self.to_v = operations.Linear(query_dim, kv_heads * dim_head, bias=bias, dtype=dtype, device=device)
+
+        self.norm_q = operations.RMSNorm(dim_head, eps=eps, dtype=dtype, device=device)
+        self.norm_k = operations.RMSNorm(dim_head, eps=eps, dtype=dtype, device=device)
+
+        self.to_out = nn.Sequential(
+            operations.Linear(heads * dim_head, query_dim, bias=bias, dtype=dtype, device=device),
+            nn.Dropout(0.0)
+        )
+
+    def forward(self, hidden_states: torch.Tensor, encoder_hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, image_rotary_emb: Optional[torch.Tensor] = None) -> torch.Tensor:
+        batch_size, sequence_length, _ = hidden_states.shape
+
+        query = self.to_q(hidden_states)
+        key = self.to_k(encoder_hidden_states)
+        value = self.to_v(encoder_hidden_states)
+
+        query = query.view(batch_size, -1, self.heads, self.dim_head)
+        key = key.view(batch_size, -1, self.kv_heads, self.dim_head)
+        value = value.view(batch_size, -1, self.kv_heads, self.dim_head)
+
+        query = self.norm_q(query)
+        key = self.norm_k(key)
+
+        if image_rotary_emb is not None:
+            query = apply_rotary_emb(query, image_rotary_emb)
+            key = apply_rotary_emb(key, image_rotary_emb)
+
+        query = query.transpose(1, 2)
+        key = key.transpose(1, 2)
+        value = value.transpose(1, 2)
+
+        if self.kv_heads < self.heads:
+            key = key.repeat_interleave(self.heads // self.kv_heads, dim=1)
+            value = value.repeat_interleave(self.heads // self.kv_heads, dim=1)
+
+        hidden_states = optimized_attention_masked(query, key, value, self.heads, attention_mask, skip_reshape=True)
+        hidden_states = self.to_out[0](hidden_states)
+        return hidden_states
+
+
+class OmniGen2TransformerBlock(nn.Module):
+    def __init__(self, dim: int, num_attention_heads: int, num_kv_heads: int, multiple_of: int, ffn_dim_multiplier: float, norm_eps: float, modulation: bool = True, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.modulation = modulation
+
+        self.attn = Attention(
+            query_dim=dim,
+            dim_head=dim // num_attention_heads,
+            heads=num_attention_heads,
+            kv_heads=num_kv_heads,
+            eps=1e-5,
+            bias=False,
+            dtype=dtype, device=device, operations=operations,
+        )
+
+        self.feed_forward = LuminaFeedForward(
+            dim=dim,
+            inner_dim=4 * dim,
+            multiple_of=multiple_of,
+            dtype=dtype, device=device, operations=operations
+        )
+
+        if modulation:
+            self.norm1 = LuminaRMSNormZero(embedding_dim=dim, norm_eps=norm_eps, dtype=dtype, device=device, operations=operations)
+        else:
+            self.norm1 = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device)
+
+        self.ffn_norm1 = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device)
+        self.norm2 = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device)
+        self.ffn_norm2 = operations.RMSNorm(dim, eps=norm_eps, dtype=dtype, device=device)
+
+    def forward(self, hidden_states: torch.Tensor, attention_mask: torch.Tensor, image_rotary_emb: torch.Tensor, temb: Optional[torch.Tensor] = None) -> torch.Tensor:
+        if self.modulation:
+            norm_hidden_states, gate_msa, scale_mlp, gate_mlp = self.norm1(hidden_states, temb)
+            attn_output = self.attn(norm_hidden_states, norm_hidden_states, attention_mask, image_rotary_emb)
+            hidden_states = hidden_states + gate_msa.unsqueeze(1).tanh() * self.norm2(attn_output)
+            mlp_output = self.feed_forward(self.ffn_norm1(hidden_states) * (1 + scale_mlp.unsqueeze(1)))
+            hidden_states = hidden_states + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2(mlp_output)
+        else:
+            norm_hidden_states = self.norm1(hidden_states)
+            attn_output = self.attn(norm_hidden_states, norm_hidden_states, attention_mask, image_rotary_emb)
+            hidden_states = hidden_states + self.norm2(attn_output)
+            mlp_output = self.feed_forward(self.ffn_norm1(hidden_states))
+            hidden_states = hidden_states + self.ffn_norm2(mlp_output)
+        return hidden_states
+
+
+class OmniGen2RotaryPosEmbed(nn.Module):
+    def __init__(self, theta: int, axes_dim: Tuple[int, int, int], axes_lens: Tuple[int, int, int] = (300, 512, 512), patch_size: int = 2):
+        super().__init__()
+        self.theta = theta
+        self.axes_dim = axes_dim
+        self.axes_lens = axes_lens
+        self.patch_size = patch_size
+        self.rope_embedder = EmbedND(dim=sum(axes_dim), theta=self.theta, axes_dim=axes_dim)
+
+    def forward(self, batch_size, encoder_seq_len, l_effective_cap_len, l_effective_ref_img_len, l_effective_img_len, ref_img_sizes, img_sizes, device):
+        p = self.patch_size
+
+        seq_lengths = [cap_len + sum(ref_img_len) + img_len for cap_len, ref_img_len, img_len in zip(l_effective_cap_len, l_effective_ref_img_len, l_effective_img_len)]
+
+        max_seq_len = max(seq_lengths)
+        max_ref_img_len = max([sum(ref_img_len) for ref_img_len in l_effective_ref_img_len])
+        max_img_len = max(l_effective_img_len)
+
+        position_ids = torch.zeros(batch_size, max_seq_len, 3, dtype=torch.int32, device=device)
+
+        for i, (cap_seq_len, seq_len) in enumerate(zip(l_effective_cap_len, seq_lengths)):
+            position_ids[i, :cap_seq_len] = repeat(torch.arange(cap_seq_len, dtype=torch.int32, device=device), "l -> l 3")
+
+            pe_shift = cap_seq_len
+            pe_shift_len = cap_seq_len
+
+            if ref_img_sizes[i] is not None:
+                for ref_img_size, ref_img_len in zip(ref_img_sizes[i], l_effective_ref_img_len[i]):
+                    H, W = ref_img_size
+                    ref_H_tokens, ref_W_tokens = H // p, W // p
+
+                    row_ids = repeat(torch.arange(ref_H_tokens, dtype=torch.int32, device=device), "h -> h w", w=ref_W_tokens).flatten()
+                    col_ids = repeat(torch.arange(ref_W_tokens, dtype=torch.int32, device=device), "w -> h w", h=ref_H_tokens).flatten()
+                    position_ids[i, pe_shift_len:pe_shift_len + ref_img_len, 0] = pe_shift
+                    position_ids[i, pe_shift_len:pe_shift_len + ref_img_len, 1] = row_ids
+                    position_ids[i, pe_shift_len:pe_shift_len + ref_img_len, 2] = col_ids
+
+                    pe_shift += max(ref_H_tokens, ref_W_tokens)
+                    pe_shift_len += ref_img_len
+
+            H, W = img_sizes[i]
+            H_tokens, W_tokens = H // p, W // p
+
+            row_ids = repeat(torch.arange(H_tokens, dtype=torch.int32, device=device), "h -> h w", w=W_tokens).flatten()
+            col_ids = repeat(torch.arange(W_tokens, dtype=torch.int32, device=device), "w -> h w", h=H_tokens).flatten()
+
+            position_ids[i, pe_shift_len: seq_len, 0] = pe_shift
+            position_ids[i, pe_shift_len: seq_len, 1] = row_ids
+            position_ids[i, pe_shift_len: seq_len, 2] = col_ids
+
+        freqs_cis = self.rope_embedder(position_ids).movedim(1, 2)
+
+        cap_freqs_cis_shape = list(freqs_cis.shape)
+        cap_freqs_cis_shape[1] = encoder_seq_len
+        cap_freqs_cis = torch.zeros(*cap_freqs_cis_shape, device=device, dtype=freqs_cis.dtype)
+
+        ref_img_freqs_cis_shape = list(freqs_cis.shape)
+        ref_img_freqs_cis_shape[1] = max_ref_img_len
+        ref_img_freqs_cis = torch.zeros(*ref_img_freqs_cis_shape, device=device, dtype=freqs_cis.dtype)
+
+        img_freqs_cis_shape = list(freqs_cis.shape)
+        img_freqs_cis_shape[1] = max_img_len
+        img_freqs_cis = torch.zeros(*img_freqs_cis_shape, device=device, dtype=freqs_cis.dtype)
+
+        for i, (cap_seq_len, ref_img_len, img_len, seq_len) in enumerate(zip(l_effective_cap_len, l_effective_ref_img_len, l_effective_img_len, seq_lengths)):
+            cap_freqs_cis[i, :cap_seq_len] = freqs_cis[i, :cap_seq_len]
+            ref_img_freqs_cis[i, :sum(ref_img_len)] = freqs_cis[i, cap_seq_len:cap_seq_len + sum(ref_img_len)]
+            img_freqs_cis[i, :img_len] = freqs_cis[i, cap_seq_len + sum(ref_img_len):cap_seq_len + sum(ref_img_len) + img_len]
+
+        return cap_freqs_cis, ref_img_freqs_cis, img_freqs_cis, freqs_cis, l_effective_cap_len, seq_lengths
+
+
+class OmniGen2Transformer2DModel(nn.Module):
+    def __init__(
+        self,
+        patch_size: int = 2,
+        in_channels: int = 16,
+        out_channels: Optional[int] = None,
+        hidden_size: int = 2304,
+        num_layers: int = 26,
+        num_refiner_layers: int = 2,
+        num_attention_heads: int = 24,
+        num_kv_heads: int = 8,
+        multiple_of: int = 256,
+        ffn_dim_multiplier: Optional[float] = None,
+        norm_eps: float = 1e-5,
+        axes_dim_rope: Tuple[int, int, int] = (32, 32, 32),
+        axes_lens: Tuple[int, int, int] = (300, 512, 512),
+        text_feat_dim: int = 1024,
+        timestep_scale: float = 1.0,
+        image_model=None,
+        device=None,
+        dtype=None,
+        operations=None,
+    ):
+        super().__init__()
+
+        self.patch_size = patch_size
+        self.out_channels = out_channels or in_channels
+        self.hidden_size = hidden_size
+        self.dtype = dtype
+
+        self.rope_embedder = OmniGen2RotaryPosEmbed(
+            theta=10000,
+            axes_dim=axes_dim_rope,
+            axes_lens=axes_lens,
+            patch_size=patch_size,
+        )
+
+        self.x_embedder = operations.Linear(patch_size * patch_size * in_channels, hidden_size, dtype=dtype, device=device)
+        self.ref_image_patch_embedder = operations.Linear(patch_size * patch_size * in_channels, hidden_size, dtype=dtype, device=device)
+
+        self.time_caption_embed = Lumina2CombinedTimestepCaptionEmbedding(
+            hidden_size=hidden_size,
+            text_feat_dim=text_feat_dim,
+            norm_eps=norm_eps,
+            timestep_scale=timestep_scale, dtype=dtype, device=device, operations=operations
+        )
+
+        self.noise_refiner = nn.ModuleList([
+            OmniGen2TransformerBlock(
+                hidden_size, num_attention_heads, num_kv_heads,
+                multiple_of, ffn_dim_multiplier, norm_eps, modulation=True, dtype=dtype, device=device, operations=operations
+            ) for _ in range(num_refiner_layers)
+        ])
+
+        self.ref_image_refiner = nn.ModuleList([
+            OmniGen2TransformerBlock(
+                hidden_size, num_attention_heads, num_kv_heads,
+                multiple_of, ffn_dim_multiplier, norm_eps, modulation=True, dtype=dtype, device=device, operations=operations
+            ) for _ in range(num_refiner_layers)
+        ])
+
+        self.context_refiner = nn.ModuleList([
+            OmniGen2TransformerBlock(
+                hidden_size, num_attention_heads, num_kv_heads,
+                multiple_of, ffn_dim_multiplier, norm_eps, modulation=False, dtype=dtype, device=device, operations=operations
+            ) for _ in range(num_refiner_layers)
+        ])
+
+        self.layers = nn.ModuleList([
+            OmniGen2TransformerBlock(
+                hidden_size, num_attention_heads, num_kv_heads,
+                multiple_of, ffn_dim_multiplier, norm_eps, modulation=True, dtype=dtype, device=device, operations=operations
+            ) for _ in range(num_layers)
+        ])
+
+        self.norm_out = LuminaLayerNormContinuous(
+            embedding_dim=hidden_size,
+            conditioning_embedding_dim=min(hidden_size, 1024),
+            elementwise_affine=False,
+            eps=1e-6,
+            out_dim=patch_size * patch_size * self.out_channels, dtype=dtype, device=device, operations=operations
+        )
+
+        self.image_index_embedding = nn.Parameter(torch.empty(5, hidden_size, device=device, dtype=dtype))
+
+    def flat_and_pad_to_seq(self, hidden_states, ref_image_hidden_states):
+        batch_size = len(hidden_states)
+        p = self.patch_size
+
+        img_sizes = [(img.size(1), img.size(2)) for img in hidden_states]
+        l_effective_img_len = [(H // p) * (W // p) for (H, W) in img_sizes]
+
+        if ref_image_hidden_states is not None:
+            ref_image_hidden_states = list(map(lambda ref: comfy.ldm.common_dit.pad_to_patch_size(ref, (p, p)), ref_image_hidden_states))
+            ref_img_sizes = [[(imgs.size(2), imgs.size(3)) if imgs is not None else None for imgs in ref_image_hidden_states]] * batch_size
+            l_effective_ref_img_len = [[(ref_img_size[0] // p) * (ref_img_size[1] // p) for ref_img_size in _ref_img_sizes] if _ref_img_sizes is not None else [0] for _ref_img_sizes in ref_img_sizes]
+        else:
+            ref_img_sizes = [None for _ in range(batch_size)]
+            l_effective_ref_img_len = [[0] for _ in range(batch_size)]
+
+        flat_ref_img_hidden_states = None
+        if ref_image_hidden_states is not None:
+            imgs = []
+            for ref_img in ref_image_hidden_states:
+                B, C, H, W = ref_img.size()
+                ref_img = rearrange(ref_img, 'b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1=p, p2=p)
+                imgs.append(ref_img)
+            flat_ref_img_hidden_states = torch.cat(imgs, dim=1)
+
+        img = hidden_states
+        B, C, H, W = img.size()
+        flat_hidden_states = rearrange(img, 'b c (h p1) (w p2) -> b (h w) (p1 p2 c)', p1=p, p2=p)
+
+        return (
+            flat_hidden_states, flat_ref_img_hidden_states,
+            None, None,
+            l_effective_ref_img_len, l_effective_img_len,
+            ref_img_sizes, img_sizes,
+        )
+
+    def img_patch_embed_and_refine(self, hidden_states, ref_image_hidden_states, padded_img_mask, padded_ref_img_mask, noise_rotary_emb, ref_img_rotary_emb, l_effective_ref_img_len, l_effective_img_len, temb):
+        batch_size = len(hidden_states)
+
+        hidden_states = self.x_embedder(hidden_states)
+        if ref_image_hidden_states is not None:
+            ref_image_hidden_states = self.ref_image_patch_embedder(ref_image_hidden_states)
+            image_index_embedding = comfy.model_management.cast_to(self.image_index_embedding, dtype=hidden_states.dtype, device=hidden_states.device)
+
+            for i in range(batch_size):
+                shift = 0
+                for j, ref_img_len in enumerate(l_effective_ref_img_len[i]):
+                    ref_image_hidden_states[i, shift:shift + ref_img_len, :] = ref_image_hidden_states[i, shift:shift + ref_img_len, :] + image_index_embedding[j]
+                    shift += ref_img_len
+
+        for layer in self.noise_refiner:
+            hidden_states = layer(hidden_states, padded_img_mask, noise_rotary_emb, temb)
+
+        if ref_image_hidden_states is not None:
+            for layer in self.ref_image_refiner:
+                ref_image_hidden_states = layer(ref_image_hidden_states, padded_ref_img_mask, ref_img_rotary_emb, temb)
+
+            hidden_states = torch.cat([ref_image_hidden_states, hidden_states], dim=1)
+
+        return hidden_states
+
+    def forward(self, x, timesteps, context, num_tokens, ref_latents=None, attention_mask=None, **kwargs):
+        B, C, H, W = x.shape
+        hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))
+        _, _, H_padded, W_padded = hidden_states.shape
+        timestep = 1.0 - timesteps
+        text_hidden_states = context
+        text_attention_mask = attention_mask
+        ref_image_hidden_states = ref_latents
+        device = hidden_states.device
+
+        temb, text_hidden_states = self.time_caption_embed(timestep, text_hidden_states, hidden_states[0].dtype)
+
+        (
+            hidden_states, ref_image_hidden_states,
+            img_mask, ref_img_mask,
+            l_effective_ref_img_len, l_effective_img_len,
+            ref_img_sizes, img_sizes,
+        ) = self.flat_and_pad_to_seq(hidden_states, ref_image_hidden_states)
+
+        (
+            context_rotary_emb, ref_img_rotary_emb, noise_rotary_emb,
+            rotary_emb, encoder_seq_lengths, seq_lengths,
+        ) = self.rope_embedder(
+            hidden_states.shape[0], text_hidden_states.shape[1], [num_tokens] * text_hidden_states.shape[0],
+            l_effective_ref_img_len, l_effective_img_len,
+            ref_img_sizes, img_sizes, device,
+        )
+
+        for layer in self.context_refiner:
+            text_hidden_states = layer(text_hidden_states, text_attention_mask, context_rotary_emb)
+
+        img_len = hidden_states.shape[1]
+        combined_img_hidden_states = self.img_patch_embed_and_refine(
+            hidden_states, ref_image_hidden_states,
+            img_mask, ref_img_mask,
+            noise_rotary_emb, ref_img_rotary_emb,
+            l_effective_ref_img_len, l_effective_img_len,
+            temb,
+        )
+
+        hidden_states = torch.cat([text_hidden_states, combined_img_hidden_states], dim=1)
+        attention_mask = None
+
+        for layer in self.layers:
+            hidden_states = layer(hidden_states, attention_mask, rotary_emb, temb)
+
+        hidden_states = self.norm_out(hidden_states, temb)
+
+        p = self.patch_size
+        output = rearrange(hidden_states[:, -img_len:], 'b (h w) (p1 p2 c) -> b c (h p1) (w p2)',  h=H_padded // p, w=W_padded// p, p1=p, p2=p)[:, :, :H, :W]
+
+        return -output

comfy/ldm/pixart/blocks.py

+# Based on:
+# https://github.com/PixArt-alpha/PixArt-alpha [Apache 2.0 license]
+# https://github.com/PixArt-alpha/PixArt-sigma [Apache 2.0 license]
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+
+from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder, Mlp, timestep_embedding
+from comfy.ldm.modules.attention import optimized_attention
+
+# if model_management.xformers_enabled():
+#     import xformers.ops
+#     if int((xformers.__version__).split(".")[2].split("+")[0]) >= 28:
+#         block_diagonal_mask_from_seqlens = xformers.ops.fmha.attn_bias.BlockDiagonalMask.from_seqlens
+#     else:
+#         block_diagonal_mask_from_seqlens = xformers.ops.fmha.BlockDiagonalMask.from_seqlens
+
+def modulate(x, shift, scale):
+    return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+
+def t2i_modulate(x, shift, scale):
+    return x * (1 + scale) + shift
+
+class MultiHeadCrossAttention(nn.Module):
+    def __init__(self, d_model, num_heads, attn_drop=0., proj_drop=0., dtype=None, device=None, operations=None, **kwargs):
+        super(MultiHeadCrossAttention, self).__init__()
+        assert d_model % num_heads == 0, "d_model must be divisible by num_heads"
+
+        self.d_model = d_model
+        self.num_heads = num_heads
+        self.head_dim = d_model // num_heads
+
+        self.q_linear = operations.Linear(d_model, d_model, dtype=dtype, device=device)
+        self.kv_linear = operations.Linear(d_model, d_model*2, dtype=dtype, device=device)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = operations.Linear(d_model, d_model, dtype=dtype, device=device)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x, cond, mask=None):
+        # query/value: img tokens; key: condition; mask: if padding tokens
+        B, N, C = x.shape
+
+        q = self.q_linear(x).view(1, -1, self.num_heads, self.head_dim)
+        kv = self.kv_linear(cond).view(1, -1, 2, self.num_heads, self.head_dim)
+        k, v = kv.unbind(2)
+
+        assert mask is None # TODO?
+        # # TODO: xformers needs separate mask logic here
+        # if model_management.xformers_enabled():
+        #     attn_bias = None
+        #     if mask is not None:
+        #         attn_bias = block_diagonal_mask_from_seqlens([N] * B, mask)
+        #     x = xformers.ops.memory_efficient_attention(q, k, v, p=0, attn_bias=attn_bias)
+        # else:
+        #     q, k, v = map(lambda t: t.transpose(1, 2), (q, k, v),)
+        #     attn_mask = None
+        #     mask = torch.ones(())
+        #     if mask is not None and len(mask) > 1:
+        #         # Create equivalent of xformer diagonal block mask, still only correct for square masks
+        #         # But depth doesn't matter as tensors can expand in that dimension
+        #         attn_mask_template = torch.ones(
+        #             [q.shape[2] // B, mask[0]],
+        #             dtype=torch.bool,
+        #             device=q.device
+        #         )
+        #         attn_mask = torch.block_diag(attn_mask_template)
+        #
+        #         # create a mask on the diagonal for each mask in the batch
+        #         for _ in range(B - 1):
+        #             attn_mask = torch.block_diag(attn_mask, attn_mask_template)
+        #     x = optimized_attention(q, k, v, self.num_heads, mask=attn_mask, skip_reshape=True)
+
+        x = optimized_attention(q.view(B, -1, C), k.view(B, -1, C), v.view(B, -1, C), self.num_heads, mask=None)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class AttentionKVCompress(nn.Module):
+    """Multi-head Attention block with KV token compression and qk norm."""
+    def __init__(self, dim, num_heads=8, qkv_bias=True, sampling='conv', sr_ratio=1, qk_norm=False, dtype=None, device=None, operations=None, **kwargs):
+        """
+        Args:
+            dim (int): Number of input channels.
+            num_heads (int): Number of attention heads.
+            qkv_bias (bool:  If True, add a learnable bias to query, key, value.
+        """
+        super().__init__()
+        assert dim % num_heads == 0, 'dim should be divisible by num_heads'
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.scale = self.head_dim ** -0.5
+
+        self.qkv = operations.Linear(dim, dim * 3, bias=qkv_bias, dtype=dtype, device=device)
+        self.proj = operations.Linear(dim, dim, dtype=dtype, device=device)
+
+        self.sampling=sampling    # ['conv', 'ave', 'uniform', 'uniform_every']
+        self.sr_ratio = sr_ratio
+        if sr_ratio > 1 and sampling == 'conv':
+            # Avg Conv Init.
+            self.sr = operations.Conv2d(dim, dim, groups=dim, kernel_size=sr_ratio, stride=sr_ratio, dtype=dtype, device=device)
+            # self.sr.weight.data.fill_(1/sr_ratio**2)
+            # self.sr.bias.data.zero_()
+            self.norm = operations.LayerNorm(dim, dtype=dtype, device=device)
+        if qk_norm:
+            self.q_norm = operations.LayerNorm(dim, dtype=dtype, device=device)
+            self.k_norm = operations.LayerNorm(dim, dtype=dtype, device=device)
+        else:
+            self.q_norm = nn.Identity()
+            self.k_norm = nn.Identity()
+
+    def downsample_2d(self, tensor, H, W, scale_factor, sampling=None):
+        if sampling is None or scale_factor == 1:
+            return tensor
+        B, N, C = tensor.shape
+
+        if sampling == 'uniform_every':
+            return tensor[:, ::scale_factor], int(N // scale_factor)
+
+        tensor = tensor.reshape(B, H, W, C).permute(0, 3, 1, 2)
+        new_H, new_W = int(H / scale_factor), int(W / scale_factor)
+        new_N = new_H * new_W
+
+        if sampling == 'ave':
+            tensor = F.interpolate(
+                tensor, scale_factor=1 / scale_factor, mode='nearest'
+            ).permute(0, 2, 3, 1)
+        elif sampling == 'uniform':
+            tensor = tensor[:, :, ::scale_factor, ::scale_factor].permute(0, 2, 3, 1)
+        elif sampling == 'conv':
+            tensor = self.sr(tensor).reshape(B, C, -1).permute(0, 2, 1)
+            tensor = self.norm(tensor)
+        else:
+            raise ValueError
+
+        return tensor.reshape(B, new_N, C).contiguous(), new_N
+
+    def forward(self, x, mask=None, HW=None, block_id=None):
+        B, N, C = x.shape # 2 4096 1152
+        new_N = N
+        if HW is None:
+            H = W = int(N ** 0.5)
+        else:
+            H, W = HW
+        qkv = self.qkv(x).reshape(B, N, 3, C)
+
+        q, k, v = qkv.unbind(2)
+        q = self.q_norm(q)
+        k = self.k_norm(k)
+
+        # KV compression
+        if self.sr_ratio > 1:
+            k, new_N = self.downsample_2d(k, H, W, self.sr_ratio, sampling=self.sampling)
+            v, new_N = self.downsample_2d(v, H, W, self.sr_ratio, sampling=self.sampling)
+
+        q = q.reshape(B, N, self.num_heads, C // self.num_heads)
+        k = k.reshape(B, new_N, self.num_heads, C // self.num_heads)
+        v = v.reshape(B, new_N, self.num_heads, C // self.num_heads)
+
+        if mask is not None:
+            raise NotImplementedError("Attn mask logic not added for self attention")
+
+        # This is never called at the moment
+        # attn_bias = None
+        # if mask is not None:
+        #     attn_bias = torch.zeros([B * self.num_heads, q.shape[1], k.shape[1]], dtype=q.dtype, device=q.device)
+        #     attn_bias.masked_fill_(mask.squeeze(1).repeat(self.num_heads, 1, 1) == 0, float('-inf'))
+
+        # attention 2
+        q, k, v = map(lambda t: t.transpose(1, 2), (q, k, v),)
+        x = optimized_attention(q, k, v, self.num_heads, mask=None, skip_reshape=True)
+
+        x = x.view(B, N, C)
+        x = self.proj(x)
+        return x
+
+
+class FinalLayer(nn.Module):
+    """
+    The final layer of PixArt.
+    """
+    def __init__(self, hidden_size, patch_size, out_channels, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm_final = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.linear = operations.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True, dtype=dtype, device=device)
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(hidden_size, 2 * hidden_size, bias=True, dtype=dtype, device=device)
+        )
+
+    def forward(self, x, c):
+        shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)
+        x = modulate(self.norm_final(x), shift, scale)
+        x = self.linear(x)
+        return x
+
+class T2IFinalLayer(nn.Module):
+    """
+    The final layer of PixArt.
+    """
+    def __init__(self, hidden_size, patch_size, out_channels, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm_final = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.linear = operations.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True, dtype=dtype, device=device)
+        self.scale_shift_table = nn.Parameter(torch.randn(2, hidden_size) / hidden_size ** 0.5)
+        self.out_channels = out_channels
+
+    def forward(self, x, t):
+        shift, scale = (self.scale_shift_table[None].to(dtype=x.dtype, device=x.device) + t[:, None]).chunk(2, dim=1)
+        x = t2i_modulate(self.norm_final(x), shift, scale)
+        x = self.linear(x)
+        return x
+
+
+class MaskFinalLayer(nn.Module):
+    """
+    The final layer of PixArt.
+    """
+    def __init__(self, final_hidden_size, c_emb_size, patch_size, out_channels, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm_final = operations.LayerNorm(final_hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.linear = operations.Linear(final_hidden_size, patch_size * patch_size * out_channels, bias=True, dtype=dtype, device=device)
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(c_emb_size, 2 * final_hidden_size, bias=True, dtype=dtype, device=device)
+        )
+    def forward(self, x, t):
+        shift, scale = self.adaLN_modulation(t).chunk(2, dim=1)
+        x = modulate(self.norm_final(x), shift, scale)
+        x = self.linear(x)
+        return x
+
+
+class DecoderLayer(nn.Module):
+    """
+    The final layer of PixArt.
+    """
+    def __init__(self, hidden_size, decoder_hidden_size, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.norm_decoder = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.linear = operations.Linear(hidden_size, decoder_hidden_size, bias=True, dtype=dtype, device=device)
+        self.adaLN_modulation = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(hidden_size, 2 * hidden_size, bias=True, dtype=dtype, device=device)
+        )
+    def forward(self, x, t):
+        shift, scale = self.adaLN_modulation(t).chunk(2, dim=1)
+        x = modulate(self.norm_decoder(x), shift, scale)
+        x = self.linear(x)
+        return x
+
+
+class SizeEmbedder(TimestepEmbedder):
+    """
+    Embeds scalar timesteps into vector representations.
+    """
+    def __init__(self, hidden_size, frequency_embedding_size=256, dtype=None, device=None, operations=None):
+        super().__init__(hidden_size=hidden_size, frequency_embedding_size=frequency_embedding_size, operations=operations)
+        self.mlp = nn.Sequential(
+            operations.Linear(frequency_embedding_size, hidden_size, bias=True, dtype=dtype, device=device),
+            nn.SiLU(),
+            operations.Linear(hidden_size, hidden_size, bias=True, dtype=dtype, device=device),
+        )
+        self.frequency_embedding_size = frequency_embedding_size
+        self.outdim = hidden_size
+
+    def forward(self, s, bs):
+        if s.ndim == 1:
+            s = s[:, None]
+        assert s.ndim == 2
+        if s.shape[0] != bs:
+            s = s.repeat(bs//s.shape[0], 1)
+            assert s.shape[0] == bs
+        b, dims = s.shape[0], s.shape[1]
+        s = rearrange(s, "b d -> (b d)")
+        s_freq = timestep_embedding(s, self.frequency_embedding_size)
+        s_emb = self.mlp(s_freq.to(s.dtype))
+        s_emb = rearrange(s_emb, "(b d) d2 -> b (d d2)", b=b, d=dims, d2=self.outdim)
+        return s_emb
+
+
+class LabelEmbedder(nn.Module):
+    """
+    Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
+    """
+    def __init__(self, num_classes, hidden_size, dropout_prob, dtype=None, device=None, operations=None):
+        super().__init__()
+        use_cfg_embedding = dropout_prob > 0
+        self.embedding_table = operations.Embedding(num_classes + use_cfg_embedding, hidden_size, dtype=dtype, device=device),
+        self.num_classes = num_classes
+        self.dropout_prob = dropout_prob
+
+    def token_drop(self, labels, force_drop_ids=None):
+        """
+        Drops labels to enable classifier-free guidance.
+        """
+        if force_drop_ids is None:
+            drop_ids = torch.rand(labels.shape[0]).cuda() < self.dropout_prob
+        else:
+            drop_ids = force_drop_ids == 1
+        labels = torch.where(drop_ids, self.num_classes, labels)
+        return labels
+
+    def forward(self, labels, train, force_drop_ids=None):
+        use_dropout = self.dropout_prob > 0
+        if (train and use_dropout) or (force_drop_ids is not None):
+            labels = self.token_drop(labels, force_drop_ids)
+        embeddings = self.embedding_table(labels)
+        return embeddings
+
+
+class CaptionEmbedder(nn.Module):
+    """
+    Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
+    """
+    def __init__(self, in_channels, hidden_size, uncond_prob, act_layer=nn.GELU(approximate='tanh'), token_num=120, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.y_proj = Mlp(
+            in_features=in_channels, hidden_features=hidden_size, out_features=hidden_size, act_layer=act_layer,
+            dtype=dtype, device=device, operations=operations,
+        )
+        self.register_buffer("y_embedding", nn.Parameter(torch.randn(token_num, in_channels) / in_channels ** 0.5))
+        self.uncond_prob = uncond_prob
+
+    def token_drop(self, caption, force_drop_ids=None):
+        """
+        Drops labels to enable classifier-free guidance.
+        """
+        if force_drop_ids is None:
+            drop_ids = torch.rand(caption.shape[0]).cuda() < self.uncond_prob
+        else:
+            drop_ids = force_drop_ids == 1
+        caption = torch.where(drop_ids[:, None, None, None], self.y_embedding, caption)
+        return caption
+
+    def forward(self, caption, train, force_drop_ids=None):
+        if train:
+            assert caption.shape[2:] == self.y_embedding.shape
+        use_dropout = self.uncond_prob > 0
+        if (train and use_dropout) or (force_drop_ids is not None):
+            caption = self.token_drop(caption, force_drop_ids)
+        caption = self.y_proj(caption)
+        return caption
+
+
+class CaptionEmbedderDoubleBr(nn.Module):
+    """
+    Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance.
+    """
+    def __init__(self, in_channels, hidden_size, uncond_prob, act_layer=nn.GELU(approximate='tanh'), token_num=120, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.proj = Mlp(
+            in_features=in_channels, hidden_features=hidden_size, out_features=hidden_size, act_layer=act_layer,
+            dtype=dtype, device=device, operations=operations,
+        )
+        self.embedding = nn.Parameter(torch.randn(1, in_channels) / 10 ** 0.5)
+        self.y_embedding = nn.Parameter(torch.randn(token_num, in_channels) / 10 ** 0.5)
+        self.uncond_prob = uncond_prob
+
+    def token_drop(self, global_caption, caption, force_drop_ids=None):
+        """
+        Drops labels to enable classifier-free guidance.
+        """
+        if force_drop_ids is None:
+            drop_ids = torch.rand(global_caption.shape[0]).cuda() < self.uncond_prob
+        else:
+            drop_ids = force_drop_ids == 1
+        global_caption = torch.where(drop_ids[:, None], self.embedding, global_caption)
+        caption = torch.where(drop_ids[:, None, None, None], self.y_embedding, caption)
+        return global_caption, caption
+
+    def forward(self, caption, train, force_drop_ids=None):
+        assert caption.shape[2: ] == self.y_embedding.shape
+        global_caption = caption.mean(dim=2).squeeze()
+        use_dropout = self.uncond_prob > 0
+        if (train and use_dropout) or (force_drop_ids is not None):
+            global_caption, caption = self.token_drop(global_caption, caption, force_drop_ids)
+        y_embed = self.proj(global_caption)
+        return y_embed, caption

comfy/ldm/pixart/pixartms.py

+# Based on:
+# https://github.com/PixArt-alpha/PixArt-alpha [Apache 2.0 license]
+# https://github.com/PixArt-alpha/PixArt-sigma [Apache 2.0 license]
+import torch
+import torch.nn as nn
+
+from .blocks import (
+    t2i_modulate,
+    CaptionEmbedder,
+    AttentionKVCompress,
+    MultiHeadCrossAttention,
+    T2IFinalLayer,
+    SizeEmbedder,
+)
+from comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder, PatchEmbed, Mlp, get_1d_sincos_pos_embed_from_grid_torch
+
+
+def get_2d_sincos_pos_embed_torch(embed_dim, w, h, pe_interpolation=1.0, base_size=16, device=None, dtype=torch.float32):
+    grid_h, grid_w = torch.meshgrid(
+        torch.arange(h, device=device, dtype=dtype) / (h/base_size) / pe_interpolation,
+        torch.arange(w, device=device, dtype=dtype) / (w/base_size) / pe_interpolation,
+        indexing='ij'
+    )
+    emb_h = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_h, device=device, dtype=dtype)
+    emb_w = get_1d_sincos_pos_embed_from_grid_torch(embed_dim // 2, grid_w, device=device, dtype=dtype)
+    emb = torch.cat([emb_w, emb_h], dim=1)  # (H*W, D)
+    return emb
+
+class PixArtMSBlock(nn.Module):
+    """
+    A PixArt block with adaptive layer norm zero (adaLN-Zero) conditioning.
+    """
+    def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, drop_path=0., input_size=None,
+                 sampling=None, sr_ratio=1, qk_norm=False, dtype=None, device=None, operations=None, **block_kwargs):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.norm1 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        self.attn = AttentionKVCompress(
+            hidden_size, num_heads=num_heads, qkv_bias=True, sampling=sampling, sr_ratio=sr_ratio,
+            qk_norm=qk_norm, dtype=dtype, device=device, operations=operations, **block_kwargs
+        )
+        self.cross_attn = MultiHeadCrossAttention(
+            hidden_size, num_heads, dtype=dtype, device=device, operations=operations, **block_kwargs
+        )
+        self.norm2 = operations.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6, dtype=dtype, device=device)
+        # to be compatible with lower version pytorch
+        approx_gelu = lambda: nn.GELU(approximate="tanh")
+        self.mlp = Mlp(
+            in_features=hidden_size, hidden_features=int(hidden_size * mlp_ratio), act_layer=approx_gelu,
+            dtype=dtype, device=device, operations=operations
+        )
+        self.scale_shift_table = nn.Parameter(torch.randn(6, hidden_size) / hidden_size ** 0.5)
+
+    def forward(self, x, y, t, mask=None, HW=None, **kwargs):
+        B, N, C = x.shape
+
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (self.scale_shift_table[None].to(dtype=x.dtype, device=x.device) + t.reshape(B, 6, -1)).chunk(6, dim=1)
+        x = x + (gate_msa * self.attn(t2i_modulate(self.norm1(x), shift_msa, scale_msa), HW=HW))
+        x = x + self.cross_attn(x, y, mask)
+        x = x + (gate_mlp * self.mlp(t2i_modulate(self.norm2(x), shift_mlp, scale_mlp)))
+
+        return x
+
+
+### Core PixArt Model ###
+class PixArtMS(nn.Module):
+    """
+    Diffusion model with a Transformer backbone.
+    """
+    def __init__(
+            self,
+            input_size=32,
+            patch_size=2,
+            in_channels=4,
+            hidden_size=1152,
+            depth=28,
+            num_heads=16,
+            mlp_ratio=4.0,
+            class_dropout_prob=0.1,
+            learn_sigma=True,
+            pred_sigma=True,
+            drop_path: float = 0.,
+            caption_channels=4096,
+            pe_interpolation=None,
+            pe_precision=None,
+            config=None,
+            model_max_length=120,
+            micro_condition=True,
+            qk_norm=False,
+            kv_compress_config=None,
+            dtype=None,
+            device=None,
+            operations=None,
+            **kwargs,
+    ):
+        nn.Module.__init__(self)
+        self.dtype = dtype
+        self.pred_sigma = pred_sigma
+        self.in_channels = in_channels
+        self.out_channels = in_channels * 2 if pred_sigma else in_channels
+        self.patch_size = patch_size
+        self.num_heads = num_heads
+        self.pe_interpolation = pe_interpolation
+        self.pe_precision = pe_precision
+        self.hidden_size = hidden_size
+        self.depth = depth
+
+        approx_gelu = lambda: nn.GELU(approximate="tanh")
+        self.t_block = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(hidden_size, 6 * hidden_size, bias=True, dtype=dtype, device=device)
+        )
+        self.x_embedder = PatchEmbed(
+            patch_size=patch_size,
+            in_chans=in_channels,
+            embed_dim=hidden_size,
+            bias=True,
+            dtype=dtype,
+            device=device,
+            operations=operations
+        )
+        self.t_embedder = TimestepEmbedder(
+            hidden_size, dtype=dtype, device=device, operations=operations,
+        )
+        self.y_embedder = CaptionEmbedder(
+            in_channels=caption_channels, hidden_size=hidden_size, uncond_prob=class_dropout_prob,
+            act_layer=approx_gelu, token_num=model_max_length,
+            dtype=dtype, device=device, operations=operations,
+        )
+
+        self.micro_conditioning = micro_condition
+        if self.micro_conditioning:
+            self.csize_embedder = SizeEmbedder(hidden_size//3, dtype=dtype, device=device, operations=operations)
+            self.ar_embedder = SizeEmbedder(hidden_size//3, dtype=dtype, device=device, operations=operations)
+
+        # For fixed sin-cos embedding:
+        # num_patches = (input_size // patch_size) * (input_size // patch_size)
+        # self.base_size = input_size // self.patch_size
+        # self.register_buffer("pos_embed", torch.zeros(1, num_patches, hidden_size))
+
+        drop_path = [x.item() for x in torch.linspace(0, drop_path, depth)]  # stochastic depth decay rule
+        if kv_compress_config is None:
+            kv_compress_config = {
+                'sampling': None,
+                'scale_factor': 1,
+                'kv_compress_layer': [],
+            }
+        self.blocks = nn.ModuleList([
+            PixArtMSBlock(
+                hidden_size, num_heads, mlp_ratio=mlp_ratio, drop_path=drop_path[i],
+                sampling=kv_compress_config['sampling'],
+                sr_ratio=int(kv_compress_config['scale_factor']) if i in kv_compress_config['kv_compress_layer'] else 1,
+                qk_norm=qk_norm,
+                dtype=dtype,
+                device=device,
+                operations=operations,
+            )
+            for i in range(depth)
+        ])
+        self.final_layer = T2IFinalLayer(
+            hidden_size, patch_size, self.out_channels, dtype=dtype, device=device, operations=operations
+        )
+
+    def forward_orig(self, x, timestep, y, mask=None, c_size=None, c_ar=None, **kwargs):
+        """
+        Original forward pass of PixArt.
+        x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
+        t: (N,) tensor of diffusion timesteps
+        y: (N, 1, 120, C) conditioning
+        ar: (N, 1): aspect ratio
+        cs: (N ,2) size conditioning for height/width
+        """
+        B, C, H, W = x.shape
+        c_res = (H + W) // 2
+        pe_interpolation = self.pe_interpolation
+        if pe_interpolation is None or self.pe_precision is not None:
+            # calculate pe_interpolation on-the-fly
+            pe_interpolation = round(c_res / (512/8.0), self.pe_precision or 0)
+
+        pos_embed = get_2d_sincos_pos_embed_torch(
+            self.hidden_size,
+            h=(H // self.patch_size),
+            w=(W // self.patch_size),
+            pe_interpolation=pe_interpolation,
+            base_size=((round(c_res / 64) * 64) // self.patch_size),
+            device=x.device,
+            dtype=x.dtype,
+        ).unsqueeze(0)
+
+        x = self.x_embedder(x) + pos_embed  # (N, T, D), where T = H * W / patch_size ** 2
+        t = self.t_embedder(timestep, x.dtype)  # (N, D)
+
+        if self.micro_conditioning and (c_size is not None and c_ar is not None):
+            bs = x.shape[0]
+            c_size = self.csize_embedder(c_size, bs)  # (N, D)
+            c_ar = self.ar_embedder(c_ar, bs)  # (N, D)
+            t = t + torch.cat([c_size, c_ar], dim=1)
+
+        t0 = self.t_block(t)
+        y = self.y_embedder(y, self.training)  # (N, D)
+
+        if mask is not None:
+            if mask.shape[0] != y.shape[0]:
+                mask = mask.repeat(y.shape[0] // mask.shape[0], 1)
+            mask = mask.squeeze(1).squeeze(1)
+            y = y.squeeze(1).masked_select(mask.unsqueeze(-1) != 0).view(1, -1, x.shape[-1])
+            y_lens = mask.sum(dim=1).tolist()
+        else:
+            y_lens = None
+            y = y.squeeze(1).view(1, -1, x.shape[-1])
+        for block in self.blocks:
+            x = block(x, y, t0, y_lens, (H, W), **kwargs)  # (N, T, D)
+
+        x = self.final_layer(x, t)  # (N, T, patch_size ** 2 * out_channels)
+        x = self.unpatchify(x, H, W)  # (N, out_channels, H, W)
+
+        return x
+
+    def forward(self, x, timesteps, context, c_size=None, c_ar=None, **kwargs):
+        B, C, H, W = x.shape
+
+        # Fallback for missing microconds
+        if self.micro_conditioning:
+            if c_size is None:
+                c_size = torch.tensor([H*8, W*8], dtype=x.dtype, device=x.device).repeat(B, 1)
+
+            if c_ar is None:
+                c_ar = torch.tensor([H/W], dtype=x.dtype, device=x.device).repeat(B, 1)
+
+        ## Still accepts the input w/o that dim but returns garbage
+        if len(context.shape) == 3:
+            context = context.unsqueeze(1)
+
+        ## run original forward pass
+        out = self.forward_orig(x, timesteps, context, c_size=c_size, c_ar=c_ar)
+
+        ## only return EPS
+        if self.pred_sigma:
+            return out[:, :self.in_channels]
+        return out
+
+    def unpatchify(self, x, h, w):
+        """
+        x: (N, T, patch_size**2 * C)
+        imgs: (N, H, W, C)
+        """
+        c = self.out_channels
+        p = self.x_embedder.patch_size[0]
+        h = h // self.patch_size
+        w = w // self.patch_size
+        assert h * w == x.shape[1]
+
+        x = x.reshape(shape=(x.shape[0], h, w, p, p, c))
+        x = torch.einsum('nhwpqc->nchpwq', x)
+        imgs = x.reshape(shape=(x.shape[0], c, h * p, w * p))
+        return imgs

comfy/ldm/qwen_image/controlnet.py

+import torch
+import math
+
+from .model import QwenImageTransformer2DModel
+
+
+class QwenImageControlNetModel(QwenImageTransformer2DModel):
+    def __init__(
+        self,
+        extra_condition_channels=0,
+        dtype=None,
+        device=None,
+        operations=None,
+        **kwargs
+    ):
+        super().__init__(final_layer=False, dtype=dtype, device=device, operations=operations, **kwargs)
+        self.main_model_double = 60
+
+        # controlnet_blocks
+        self.controlnet_blocks = torch.nn.ModuleList([])
+        for _ in range(len(self.transformer_blocks)):
+            self.controlnet_blocks.append(operations.Linear(self.inner_dim, self.inner_dim, device=device, dtype=dtype))
+        self.controlnet_x_embedder = operations.Linear(self.in_channels + extra_condition_channels, self.inner_dim, device=device, dtype=dtype)
+
+    def forward(
+        self,
+        x,
+        timesteps,
+        context,
+        attention_mask=None,
+        guidance: torch.Tensor = None,
+        ref_latents=None,
+        hint=None,
+        transformer_options={},
+        **kwargs
+    ):
+        timestep = timesteps
+        encoder_hidden_states = context
+        encoder_hidden_states_mask = attention_mask
+
+        hidden_states, img_ids, orig_shape = self.process_img(x)
+        hint, _, _ = self.process_img(hint)
+
+        txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2))
+        txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)
+        del ids, txt_ids, img_ids
+
+        hidden_states = self.img_in(hidden_states) + self.controlnet_x_embedder(hint)
+        encoder_hidden_states = self.txt_norm(encoder_hidden_states)
+        encoder_hidden_states = self.txt_in(encoder_hidden_states)
+
+        if guidance is not None:
+            guidance = guidance * 1000
+
+        temb = (
+            self.time_text_embed(timestep, hidden_states)
+            if guidance is None
+            else self.time_text_embed(timestep, guidance, hidden_states)
+        )
+
+        repeat = math.ceil(self.main_model_double / len(self.controlnet_blocks))
+
+        controlnet_block_samples = ()
+        for i, block in enumerate(self.transformer_blocks):
+            encoder_hidden_states, hidden_states = block(
+                hidden_states=hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_hidden_states_mask=encoder_hidden_states_mask,
+                temb=temb,
+                image_rotary_emb=image_rotary_emb,
+            )
+
+            controlnet_block_samples = controlnet_block_samples + (self.controlnet_blocks[i](hidden_states),) * repeat
+
+        return {"input": controlnet_block_samples[:self.main_model_double]}

comfy/ldm/qwen_image/model.py

+# https://github.com/QwenLM/Qwen-Image (Apache 2.0)
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Optional, Tuple
+from einops import repeat
+
+from comfy.ldm.lightricks.model import TimestepEmbedding, Timesteps
+from comfy.ldm.modules.attention import optimized_attention_masked
+from comfy.ldm.flux.layers import EmbedND
+import comfy.ldm.common_dit
+import comfy.patcher_extension
+
+class GELU(nn.Module):
+    def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool = True, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.proj = operations.Linear(dim_in, dim_out, bias=bias, dtype=dtype, device=device)
+        self.approximate = approximate
+
+    def forward(self, hidden_states):
+        hidden_states = self.proj(hidden_states)
+        hidden_states = F.gelu(hidden_states, approximate=self.approximate)
+        return hidden_states
+
+
+class FeedForward(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        dim_out: Optional[int] = None,
+        mult: int = 4,
+        dropout: float = 0.0,
+        inner_dim=None,
+        bias: bool = True,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        if inner_dim is None:
+            inner_dim = int(dim * mult)
+        dim_out = dim_out if dim_out is not None else dim
+
+        self.net = nn.ModuleList([])
+        self.net.append(GELU(dim, inner_dim, approximate="tanh", bias=bias, dtype=dtype, device=device, operations=operations))
+        self.net.append(nn.Dropout(dropout))
+        self.net.append(operations.Linear(inner_dim, dim_out, bias=bias, dtype=dtype, device=device))
+
+    def forward(self, hidden_states: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+        for module in self.net:
+            hidden_states = module(hidden_states)
+        return hidden_states
+
+
+def apply_rotary_emb(x, freqs_cis):
+    if x.shape[1] == 0:
+        return x
+
+    t_ = x.reshape(*x.shape[:-1], -1, 1, 2)
+    t_out = freqs_cis[..., 0] * t_[..., 0] + freqs_cis[..., 1] * t_[..., 1]
+    return t_out.reshape(*x.shape)
+
+
+class QwenTimestepProjEmbeddings(nn.Module):
+    def __init__(self, embedding_dim, pooled_projection_dim, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0, scale=1000)
+        self.timestep_embedder = TimestepEmbedding(
+            in_channels=256,
+            time_embed_dim=embedding_dim,
+            dtype=dtype,
+            device=device,
+            operations=operations
+        )
+
+    def forward(self, timestep, hidden_states):
+        timesteps_proj = self.time_proj(timestep)
+        timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_states.dtype))
+        return timesteps_emb
+
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        query_dim: int,
+        dim_head: int = 64,
+        heads: int = 8,
+        dropout: float = 0.0,
+        bias: bool = False,
+        eps: float = 1e-5,
+        out_bias: bool = True,
+        out_dim: int = None,
+        out_context_dim: int = None,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.inner_dim = out_dim if out_dim is not None else dim_head * heads
+        self.inner_kv_dim = self.inner_dim
+        self.heads = heads
+        self.dim_head = dim_head
+        self.out_dim = out_dim if out_dim is not None else query_dim
+        self.out_context_dim = out_context_dim if out_context_dim is not None else query_dim
+        self.dropout = dropout
+
+        # Q/K normalization
+        self.norm_q = operations.RMSNorm(dim_head, eps=eps, elementwise_affine=True, dtype=dtype, device=device)
+        self.norm_k = operations.RMSNorm(dim_head, eps=eps, elementwise_affine=True, dtype=dtype, device=device)
+        self.norm_added_q = operations.RMSNorm(dim_head, eps=eps, dtype=dtype, device=device)
+        self.norm_added_k = operations.RMSNorm(dim_head, eps=eps, dtype=dtype, device=device)
+
+        # Image stream projections
+        self.to_q = operations.Linear(query_dim, self.inner_dim, bias=bias, dtype=dtype, device=device)
+        self.to_k = operations.Linear(query_dim, self.inner_kv_dim, bias=bias, dtype=dtype, device=device)
+        self.to_v = operations.Linear(query_dim, self.inner_kv_dim, bias=bias, dtype=dtype, device=device)
+
+        # Text stream projections
+        self.add_q_proj = operations.Linear(query_dim, self.inner_dim, bias=bias, dtype=dtype, device=device)
+        self.add_k_proj = operations.Linear(query_dim, self.inner_kv_dim, bias=bias, dtype=dtype, device=device)
+        self.add_v_proj = operations.Linear(query_dim, self.inner_kv_dim, bias=bias, dtype=dtype, device=device)
+
+        # Output projections
+        self.to_out = nn.ModuleList([
+            operations.Linear(self.inner_dim, self.out_dim, bias=out_bias, dtype=dtype, device=device),
+            nn.Dropout(dropout)
+        ])
+        self.to_add_out = operations.Linear(self.inner_dim, self.out_context_dim, bias=out_bias, dtype=dtype, device=device)
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,  # Image stream
+        encoder_hidden_states: torch.FloatTensor = None,  # Text stream
+        encoder_hidden_states_mask: torch.FloatTensor = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        image_rotary_emb: Optional[torch.Tensor] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        seq_txt = encoder_hidden_states.shape[1]
+
+        img_query = self.to_q(hidden_states).unflatten(-1, (self.heads, -1))
+        img_key = self.to_k(hidden_states).unflatten(-1, (self.heads, -1))
+        img_value = self.to_v(hidden_states).unflatten(-1, (self.heads, -1))
+
+        txt_query = self.add_q_proj(encoder_hidden_states).unflatten(-1, (self.heads, -1))
+        txt_key = self.add_k_proj(encoder_hidden_states).unflatten(-1, (self.heads, -1))
+        txt_value = self.add_v_proj(encoder_hidden_states).unflatten(-1, (self.heads, -1))
+
+        img_query = self.norm_q(img_query)
+        img_key = self.norm_k(img_key)
+        txt_query = self.norm_added_q(txt_query)
+        txt_key = self.norm_added_k(txt_key)
+
+        joint_query = torch.cat([txt_query, img_query], dim=1)
+        joint_key = torch.cat([txt_key, img_key], dim=1)
+        joint_value = torch.cat([txt_value, img_value], dim=1)
+
+        joint_query = apply_rotary_emb(joint_query, image_rotary_emb)
+        joint_key = apply_rotary_emb(joint_key, image_rotary_emb)
+
+        joint_query = joint_query.flatten(start_dim=2)
+        joint_key = joint_key.flatten(start_dim=2)
+        joint_value = joint_value.flatten(start_dim=2)
+
+        joint_hidden_states = optimized_attention_masked(joint_query, joint_key, joint_value, self.heads, attention_mask)
+
+        txt_attn_output = joint_hidden_states[:, :seq_txt, :]
+        img_attn_output = joint_hidden_states[:, seq_txt:, :]
+
+        img_attn_output = self.to_out[0](img_attn_output)
+        img_attn_output = self.to_out[1](img_attn_output)
+        txt_attn_output = self.to_add_out(txt_attn_output)
+
+        return img_attn_output, txt_attn_output
+
+
+class QwenImageTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        eps: float = 1e-6,
+        dtype=None,
+        device=None,
+        operations=None
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+
+        self.img_mod = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(dim, 6 * dim, bias=True, dtype=dtype, device=device),
+        )
+        self.img_norm1 = operations.LayerNorm(dim, elementwise_affine=False, eps=eps, dtype=dtype, device=device)
+        self.img_norm2 = operations.LayerNorm(dim, elementwise_affine=False, eps=eps, dtype=dtype, device=device)
+        self.img_mlp = FeedForward(dim=dim, dim_out=dim, dtype=dtype, device=device, operations=operations)
+
+        self.txt_mod = nn.Sequential(
+            nn.SiLU(),
+            operations.Linear(dim, 6 * dim, bias=True, dtype=dtype, device=device),
+        )
+        self.txt_norm1 = operations.LayerNorm(dim, elementwise_affine=False, eps=eps, dtype=dtype, device=device)
+        self.txt_norm2 = operations.LayerNorm(dim, elementwise_affine=False, eps=eps, dtype=dtype, device=device)
+        self.txt_mlp = FeedForward(dim=dim, dim_out=dim, dtype=dtype, device=device, operations=operations)
+
+        self.attn = Attention(
+            query_dim=dim,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=dim,
+            bias=True,
+            eps=eps,
+            dtype=dtype,
+            device=device,
+            operations=operations,
+        )
+
+    def _modulate(self, x: torch.Tensor, mod_params: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+        shift, scale, gate = torch.chunk(mod_params, 3, dim=-1)
+        return torch.addcmul(shift.unsqueeze(1), x, 1 + scale.unsqueeze(1)), gate.unsqueeze(1)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: torch.Tensor,
+        encoder_hidden_states_mask: torch.Tensor,
+        temb: torch.Tensor,
+        image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        img_mod_params = self.img_mod(temb)
+        txt_mod_params = self.txt_mod(temb)
+        img_mod1, img_mod2 = img_mod_params.chunk(2, dim=-1)
+        txt_mod1, txt_mod2 = txt_mod_params.chunk(2, dim=-1)
+
+        img_normed = self.img_norm1(hidden_states)
+        img_modulated, img_gate1 = self._modulate(img_normed, img_mod1)
+        txt_normed = self.txt_norm1(encoder_hidden_states)
+        txt_modulated, txt_gate1 = self._modulate(txt_normed, txt_mod1)
+
+        img_attn_output, txt_attn_output = self.attn(
+            hidden_states=img_modulated,
+            encoder_hidden_states=txt_modulated,
+            encoder_hidden_states_mask=encoder_hidden_states_mask,
+            image_rotary_emb=image_rotary_emb,
+        )
+
+        hidden_states = hidden_states + img_gate1 * img_attn_output
+        encoder_hidden_states = encoder_hidden_states + txt_gate1 * txt_attn_output
+
+        img_normed2 = self.img_norm2(hidden_states)
+        img_modulated2, img_gate2 = self._modulate(img_normed2, img_mod2)
+        hidden_states = torch.addcmul(hidden_states, img_gate2, self.img_mlp(img_modulated2))
+
+        txt_normed2 = self.txt_norm2(encoder_hidden_states)
+        txt_modulated2, txt_gate2 = self._modulate(txt_normed2, txt_mod2)
+        encoder_hidden_states = torch.addcmul(encoder_hidden_states, txt_gate2, self.txt_mlp(txt_modulated2))
+
+        return encoder_hidden_states, hidden_states
+
+
+class LastLayer(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        conditioning_embedding_dim: int,
+        elementwise_affine=False,
+        eps=1e-6,
+        bias=True,
+        dtype=None, device=None, operations=None
+    ):
+        super().__init__()
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(conditioning_embedding_dim, embedding_dim * 2, bias=bias, dtype=dtype, device=device)
+        self.norm = operations.LayerNorm(embedding_dim, eps, elementwise_affine=False, bias=bias, dtype=dtype, device=device)
+
+    def forward(self, x: torch.Tensor, conditioning_embedding: torch.Tensor) -> torch.Tensor:
+        emb = self.linear(self.silu(conditioning_embedding))
+        scale, shift = torch.chunk(emb, 2, dim=1)
+        x = torch.addcmul(shift[:, None, :], self.norm(x), (1 + scale)[:, None, :])
+        return x
+
+
+class QwenImageTransformer2DModel(nn.Module):
+    def __init__(
+        self,
+        patch_size: int = 2,
+        in_channels: int = 64,
+        out_channels: Optional[int] = 16,
+        num_layers: int = 60,
+        attention_head_dim: int = 128,
+        num_attention_heads: int = 24,
+        joint_attention_dim: int = 3584,
+        pooled_projection_dim: int = 768,
+        guidance_embeds: bool = False,
+        axes_dims_rope: Tuple[int, int, int] = (16, 56, 56),
+        image_model=None,
+        final_layer=True,
+        dtype=None,
+        device=None,
+        operations=None,
+    ):
+        super().__init__()
+        self.dtype = dtype
+        self.patch_size = patch_size
+        self.in_channels = in_channels
+        self.out_channels = out_channels or in_channels
+        self.inner_dim = num_attention_heads * attention_head_dim
+
+        self.pe_embedder = EmbedND(dim=attention_head_dim, theta=10000, axes_dim=list(axes_dims_rope))
+
+        self.time_text_embed = QwenTimestepProjEmbeddings(
+            embedding_dim=self.inner_dim,
+            pooled_projection_dim=pooled_projection_dim,
+            dtype=dtype,
+            device=device,
+            operations=operations
+        )
+
+        self.txt_norm = operations.RMSNorm(joint_attention_dim, eps=1e-6, dtype=dtype, device=device)
+        self.img_in = operations.Linear(in_channels, self.inner_dim, dtype=dtype, device=device)
+        self.txt_in = operations.Linear(joint_attention_dim, self.inner_dim, dtype=dtype, device=device)
+
+        self.transformer_blocks = nn.ModuleList([
+            QwenImageTransformerBlock(
+                dim=self.inner_dim,
+                num_attention_heads=num_attention_heads,
+                attention_head_dim=attention_head_dim,
+                dtype=dtype,
+                device=device,
+                operations=operations
+            )
+            for _ in range(num_layers)
+        ])
+
+        if final_layer:
+            self.norm_out = LastLayer(self.inner_dim, self.inner_dim, dtype=dtype, device=device, operations=operations)
+            self.proj_out = operations.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True, dtype=dtype, device=device)
+
+    def process_img(self, x, index=0, h_offset=0, w_offset=0):
+        bs, c, t, h, w = x.shape
+        patch_size = self.patch_size
+        hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (1, self.patch_size, self.patch_size))
+        orig_shape = hidden_states.shape
+        hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2)
+        hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5)
+        hidden_states = hidden_states.reshape(orig_shape[0], (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4)
+        h_len = ((h + (patch_size // 2)) // patch_size)
+        w_len = ((w + (patch_size // 2)) // patch_size)
+
+        h_offset = ((h_offset + (patch_size // 2)) // patch_size)
+        w_offset = ((w_offset + (patch_size // 2)) // patch_size)
+
+        img_ids = torch.zeros((h_len, w_len, 3), device=x.device)
+        img_ids[:, :, 0] = img_ids[:, :, 1] + index
+        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1) - (h_len // 2)
+        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0) - (w_len // 2)
+        return hidden_states, repeat(img_ids, "h w c -> b (h w) c", b=bs), orig_shape
+
+    def forward(self, x, timestep, context, attention_mask=None, guidance=None, ref_latents=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, attention_mask, guidance, ref_latents, transformer_options, **kwargs)
+
+    def _forward(
+        self,
+        x,
+        timesteps,
+        context,
+        attention_mask=None,
+        guidance: torch.Tensor = None,
+        ref_latents=None,
+        transformer_options={},
+        control=None,
+        **kwargs
+    ):
+        timestep = timesteps
+        encoder_hidden_states = context
+        encoder_hidden_states_mask = attention_mask
+
+        hidden_states, img_ids, orig_shape = self.process_img(x)
+        num_embeds = hidden_states.shape[1]
+
+        if ref_latents is not None:
+            h = 0
+            w = 0
+            index = 0
+            index_ref_method = kwargs.get("ref_latents_method", "index") == "index"
+            for ref in ref_latents:
+                if index_ref_method:
+                    index += 1
+                    h_offset = 0
+                    w_offset = 0
+                else:
+                    index = 1
+                    h_offset = 0
+                    w_offset = 0
+                    if ref.shape[-2] + h > ref.shape[-1] + w:
+                        w_offset = w
+                    else:
+                        h_offset = h
+                    h = max(h, ref.shape[-2] + h_offset)
+                    w = max(w, ref.shape[-1] + w_offset)
+
+                kontext, kontext_ids, _ = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset)
+                hidden_states = torch.cat([hidden_states, kontext], dim=1)
+                img_ids = torch.cat([img_ids, kontext_ids], dim=1)
+
+        txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2))
+        txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)
+        del ids, txt_ids, img_ids
+
+        hidden_states = self.img_in(hidden_states)
+        encoder_hidden_states = self.txt_norm(encoder_hidden_states)
+        encoder_hidden_states = self.txt_in(encoder_hidden_states)
+
+        if guidance is not None:
+            guidance = guidance * 1000
+
+        temb = (
+            self.time_text_embed(timestep, hidden_states)
+            if guidance is None
+            else self.time_text_embed(timestep, guidance, hidden_states)
+        )
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        patches = transformer_options.get("patches", {})
+        blocks_replace = patches_replace.get("dit", {})
+
+        for i, block in enumerate(self.transformer_blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["txt"], out["img"] = block(hidden_states=args["img"], encoder_hidden_states=args["txt"], encoder_hidden_states_mask=encoder_hidden_states_mask, temb=args["vec"], image_rotary_emb=args["pe"])
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": hidden_states, "txt": encoder_hidden_states, "vec": temb, "pe": image_rotary_emb}, {"original_block": block_wrap})
+                hidden_states = out["img"]
+                encoder_hidden_states = out["txt"]
+            else:
+                encoder_hidden_states, hidden_states = block(
+                    hidden_states=hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_hidden_states_mask=encoder_hidden_states_mask,
+                    temb=temb,
+                    image_rotary_emb=image_rotary_emb,
+                )
+
+            if "double_block" in patches:
+                for p in patches["double_block"]:
+                    out = p({"img": hidden_states, "txt": encoder_hidden_states, "x": x, "block_index": i})
+                    hidden_states = out["img"]
+                    encoder_hidden_states = out["txt"]
+
+            if control is not None: # Controlnet
+                control_i = control.get("input")
+                if i < len(control_i):
+                    add = control_i[i]
+                    if add is not None:
+                        hidden_states[:, :add.shape[1]] += add
+
+        hidden_states = self.norm_out(hidden_states, temb)
+        hidden_states = self.proj_out(hidden_states)
+
+        hidden_states = hidden_states[:, :num_embeds].view(orig_shape[0], orig_shape[-2] // 2, orig_shape[-1] // 2, orig_shape[1], 2, 2)
+        hidden_states = hidden_states.permute(0, 3, 1, 4, 2, 5)
+        return hidden_states.reshape(orig_shape)[:, :, :, :x.shape[-2], :x.shape[-1]]

comfy/ldm/util.py

+import importlib
+import logging
+
+import torch
+from torch import optim
+import numpy as np
+
+from inspect import isfunction
+from PIL import Image, ImageDraw, ImageFont
+
+
+def log_txt_as_img(wh, xc, size=10):
+    # wh a tuple of (width, height)
+    # xc a list of captions to plot
+    b = len(xc)
+    txts = list()
+    for bi in range(b):
+        txt = Image.new("RGB", wh, color="white")
+        draw = ImageDraw.Draw(txt)
+        font = ImageFont.truetype('data/DejaVuSans.ttf', size=size)
+        nc = int(40 * (wh[0] / 256))
+        lines = "\n".join(xc[bi][start:start + nc] for start in range(0, len(xc[bi]), nc))
+
+        try:
+            draw.text((0, 0), lines, fill="black", font=font)
+        except UnicodeEncodeError:
+            logging.warning("Cant encode string for logging. Skipping.")
+
+        txt = np.array(txt).transpose(2, 0, 1) / 127.5 - 1.0
+        txts.append(txt)
+    txts = np.stack(txts)
+    txts = torch.tensor(txts)
+    return txts
+
+
+def ismap(x):
+    if not isinstance(x, torch.Tensor):
+        return False
+    return (len(x.shape) == 4) and (x.shape[1] > 3)
+
+
+def isimage(x):
+    if not isinstance(x,torch.Tensor):
+        return False
+    return (len(x.shape) == 4) and (x.shape[1] == 3 or x.shape[1] == 1)
+
+
+def exists(x):
+    return x is not None
+
+
+def default(val, d):
+    if exists(val):
+        return val
+    return d() if isfunction(d) else d
+
+
+def mean_flat(tensor):
+    """
+    https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/nn.py#L86
+    Take the mean over all non-batch dimensions.
+    """
+    return tensor.mean(dim=list(range(1, len(tensor.shape))))
+
+
+def count_params(model, verbose=False):
+    total_params = sum(p.numel() for p in model.parameters())
+    if verbose:
+        logging.info(f"{model.__class__.__name__} has {total_params*1.e-6:.2f} M params.")
+    return total_params
+
+
+def instantiate_from_config(config):
+    if not "target" in config:
+        if config == '__is_first_stage__':
+            return None
+        elif config == "__is_unconditional__":
+            return None
+        raise KeyError("Expected key `target` to instantiate.")
+    return get_obj_from_str(config["target"])(**config.get("params", dict()))
+
+
+def get_obj_from_str(string, reload=False):
+    module, cls = string.rsplit(".", 1)
+    if reload:
+        module_imp = importlib.import_module(module)
+        importlib.reload(module_imp)
+    return getattr(importlib.import_module(module, package=None), cls)
+
+
+class AdamWwithEMAandWings(optim.Optimizer):
+    # credit to https://gist.github.com/crowsonkb/65f7265353f403714fce3b2595e0b298
+    def __init__(self, params, lr=1.e-3, betas=(0.9, 0.999), eps=1.e-8,  # TODO: check hyperparameters before using
+                 weight_decay=1.e-2, amsgrad=False, ema_decay=0.9999,   # ema decay to match previous code
+                 ema_power=1., param_names=()):
+        """AdamW that saves EMA versions of the parameters."""
+        if not 0.0 <= lr:
+            raise ValueError("Invalid learning rate: {}".format(lr))
+        if not 0.0 <= eps:
+            raise ValueError("Invalid epsilon value: {}".format(eps))
+        if not 0.0 <= betas[0] < 1.0:
+            raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
+        if not 0.0 <= betas[1] < 1.0:
+            raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
+        if not 0.0 <= weight_decay:
+            raise ValueError("Invalid weight_decay value: {}".format(weight_decay))
+        if not 0.0 <= ema_decay <= 1.0:
+            raise ValueError("Invalid ema_decay value: {}".format(ema_decay))
+        defaults = dict(lr=lr, betas=betas, eps=eps,
+                        weight_decay=weight_decay, amsgrad=amsgrad, ema_decay=ema_decay,
+                        ema_power=ema_power, param_names=param_names)
+        super().__init__(params, defaults)
+
+    def __setstate__(self, state):
+        super().__setstate__(state)
+        for group in self.param_groups:
+            group.setdefault('amsgrad', False)
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step.
+        Args:
+            closure (callable, optional): A closure that reevaluates the model
+                and returns the loss.
+        """
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            params_with_grad = []
+            grads = []
+            exp_avgs = []
+            exp_avg_sqs = []
+            ema_params_with_grad = []
+            max_exp_avg_sqs = []
+            state_steps = []
+            amsgrad = group['amsgrad']
+            beta1, beta2 = group['betas']
+            ema_decay = group['ema_decay']
+            ema_power = group['ema_power']
+
+            for p in group['params']:
+                if p.grad is None:
+                    continue
+                params_with_grad.append(p)
+                if p.grad.is_sparse:
+                    raise RuntimeError('AdamW does not support sparse gradients')
+                grads.append(p.grad)
+
+                state = self.state[p]
+
+                # State initialization
+                if len(state) == 0:
+                    state['step'] = 0
+                    # Exponential moving average of gradient values
+                    state['exp_avg'] = torch.zeros_like(p, memory_format=torch.preserve_format)
+                    # Exponential moving average of squared gradient values
+                    state['exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
+                    if amsgrad:
+                        # Maintains max of all exp. moving avg. of sq. grad. values
+                        state['max_exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)
+                    # Exponential moving average of parameter values
+                    state['param_exp_avg'] = p.detach().float().clone()
+
+                exp_avgs.append(state['exp_avg'])
+                exp_avg_sqs.append(state['exp_avg_sq'])
+                ema_params_with_grad.append(state['param_exp_avg'])
+
+                if amsgrad:
+                    max_exp_avg_sqs.append(state['max_exp_avg_sq'])
+
+                # update the steps for each param group update
+                state['step'] += 1
+                # record the step after step update
+                state_steps.append(state['step'])
+
+            optim._functional.adamw(params_with_grad,
+                    grads,
+                    exp_avgs,
+                    exp_avg_sqs,
+                    max_exp_avg_sqs,
+                    state_steps,
+                    amsgrad=amsgrad,
+                    beta1=beta1,
+                    beta2=beta2,
+                    lr=group['lr'],
+                    weight_decay=group['weight_decay'],
+                    eps=group['eps'],
+                    maximize=False)
+
+            cur_ema_decay = min(ema_decay, 1 - state['step'] ** -ema_power)
+            for param, ema_param in zip(params_with_grad, ema_params_with_grad):
+                ema_param.mul_(cur_ema_decay).add_(param.float(), alpha=1 - cur_ema_decay)
+
+        return loss

comfy/ldm/wan/model.py

+# original version: https://github.com/Wan-Video/Wan2.1/blob/main/wan/modules/model.py
+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+import math
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+
+from comfy.ldm.modules.attention import optimized_attention
+from comfy.ldm.flux.layers import EmbedND
+from comfy.ldm.flux.math import apply_rope
+import comfy.ldm.common_dit
+import comfy.model_management
+import comfy.patcher_extension
+from lightop import op, LayerNorm, RMSNorm
+
+def sinusoidal_embedding_1d(dim, position):
+    # preprocess
+    assert dim % 2 == 0
+    half = dim // 2
+    position = position.type(torch.float32)
+
+    # calculation
+    sinusoid = torch.outer(
+        position, torch.pow(10000, -torch.arange(half).to(position).div(half)))
+    x = torch.cat([torch.cos(sinusoid), torch.sin(sinusoid)], dim=1)
+    return x
+
+
+class WanSelfAttention(nn.Module):
+
+    def __init__(self,
+                 dim,
+                 num_heads,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 eps=1e-6, operation_settings={}):
+        assert dim % num_heads == 0
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.window_size = window_size
+        self.qk_norm = qk_norm
+        self.eps = eps
+
+        # layers
+        self.q = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.k = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.v = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.o = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        # self.norm_q = operation_settings.get("operations").RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
+        # self.norm_k = operation_settings.get("operations").RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
+        self.norm_q = RMSNorm(dim, eps=eps, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
+        self.norm_k = RMSNorm(dim, eps=eps, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
+        
+
+    def forward(self, x, freqs):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L, num_heads, C / num_heads]
+            freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
+        """
+        b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
+
+        # query, key, value function
+        def qkv_fn(x):
+            q = self.norm_q(self.q(x)).view(b, s, n, d)
+            k = self.norm_k(self.k(x)).view(b, s, n, d)
+            v = self.v(x).view(b, s, n * d)
+            return q, k, v
+
+        q, k, v = qkv_fn(x)
+        q, k = apply_rope(q, k, freqs)
+        # import pdb
+        # pdb.set_trace()
+        x = optimized_attention(
+            q.view(b, s, n * d),
+            k.view(b, s, n * d),
+            v,
+            heads=self.num_heads,
+        )
+
+        x = self.o(x)
+        return x
+
+
+class WanT2VCrossAttention(WanSelfAttention):
+
+    def forward(self, x, context, **kwargs):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L1, C]
+            context(Tensor): Shape [B, L2, C]
+        """
+        # compute query, key, value
+        q = self.norm_q(self.q(x))
+        k = self.norm_k(self.k(context))
+        v = self.v(context)
+
+        # compute attention
+        x = optimized_attention(q, k, v, heads=self.num_heads)
+
+        x = self.o(x)
+        return x
+
+
+class WanI2VCrossAttention(WanSelfAttention):
+
+    def __init__(self,
+                 dim,
+                 num_heads,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 eps=1e-6, operation_settings={}):
+        super().__init__(dim, num_heads, window_size, qk_norm, eps, operation_settings=operation_settings)
+
+        self.k_img = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.v_img = operation_settings.get("operations").Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        # self.alpha = nn.Parameter(torch.zeros((1, )))
+        self.norm_k_img = operation_settings.get("operations").RMSNorm(dim, eps=eps, elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if qk_norm else nn.Identity()
+
+    def forward(self, x, context, context_img_len):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L1, C]
+            context(Tensor): Shape [B, L2, C]
+        """
+        context_img = context[:, :context_img_len]
+        context = context[:, context_img_len:]
+
+        # compute query, key, value
+        q = self.norm_q(self.q(x))
+        k = self.norm_k(self.k(context))
+        v = self.v(context)
+        k_img = self.norm_k_img(self.k_img(context_img))
+        v_img = self.v_img(context_img)
+        img_x = optimized_attention(q, k_img, v_img, heads=self.num_heads)
+        # compute attention
+        x = optimized_attention(q, k, v, heads=self.num_heads)
+
+        # output
+        x = x + img_x
+        x = self.o(x)
+        return x
+
+
+WAN_CROSSATTENTION_CLASSES = {
+    't2v_cross_attn': WanT2VCrossAttention,
+    'i2v_cross_attn': WanI2VCrossAttention,
+}
+
+
+def repeat_e(e, x):
+    repeats = 1
+    if e.size(1) > 1:
+        repeats = x.size(1) // e.size(1)
+    if repeats == 1:
+        return e
+    if repeats * e.size(1) == x.size(1):
+        return torch.repeat_interleave(e, repeats, dim=1)
+    else:
+        return torch.repeat_interleave(e, repeats + 1, dim=1)[:, :x.size(1)]
+
+
+class WanAttentionBlock(nn.Module):
+
+    def __init__(self,
+                 cross_attn_type,
+                 dim,
+                 ffn_dim,
+                 num_heads,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=False,
+                 eps=1e-6, operation_settings={}):
+        super().__init__()
+        self.dim = dim
+        self.ffn_dim = ffn_dim
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.qk_norm = qk_norm
+        self.cross_attn_norm = cross_attn_norm
+        self.eps = eps
+
+        # layers
+        self.norm1 = operation_settings.get("operations").LayerNorm(dim, eps, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        # self.norm1 = LayerNorm(dim, eps, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.self_attn = WanSelfAttention(dim, num_heads, window_size, qk_norm,
+                                          eps, operation_settings=operation_settings)
+        self.norm3 = operation_settings.get("operations").LayerNorm(
+            dim, eps,
+            elementwise_affine=True, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if cross_attn_norm else nn.Identity()
+        # self.norm3 = LayerNorm(
+        #     dim, eps,
+        #     device=operation_settings.get("device"), dtype=operation_settings.get("dtype")) if cross_attn_norm else nn.Identity()
+        self.cross_attn = WAN_CROSSATTENTION_CLASSES[cross_attn_type](dim,
+                                                                      num_heads,
+                                                                      (-1, -1),
+                                                                      qk_norm,
+                                                                      eps, operation_settings=operation_settings)
+        self.norm2 = operation_settings.get("operations").LayerNorm(dim, eps, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        # self.norm2 = LayerNorm(dim, eps, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.ffn = nn.Sequential(
+            operation_settings.get("operations").Linear(dim, ffn_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), nn.GELU(approximate='tanh'),
+            operation_settings.get("operations").Linear(ffn_dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+        # modulation
+        self.modulation = nn.Parameter(torch.empty(1, 6, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+    def forward(
+        self,
+        x,
+        e,
+        freqs,
+        context,
+        context_img_len=257,
+    ):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L, C]
+            e(Tensor): Shape [B, 6, C]
+            freqs(Tensor): Rope freqs, shape [1024, C / num_heads / 2]
+        """
+        # assert e.dtype == torch.float32
+
+        if e.ndim < 4:
+            e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e).chunk(6, dim=1)
+        else:
+            e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device).unsqueeze(0) + e).unbind(2)
+        # assert e[0].dtype == torch.float32
+
+        # self-attention
+        y = self.self_attn(
+            torch.addcmul(repeat_e(e[0], x), self.norm1(x), 1 + repeat_e(e[1], x)),
+            freqs)
+
+        x = torch.addcmul(x, y, repeat_e(e[2], x))
+
+        # cross-attention & ffn
+        x = x + self.cross_attn(self.norm3(x), context, context_img_len=context_img_len)
+        y = self.ffn(torch.addcmul(repeat_e(e[3], x), self.norm2(x), 1 + repeat_e(e[4], x)))
+        x = torch.addcmul(x, y, repeat_e(e[5], x))
+        return x
+
+
+class VaceWanAttentionBlock(WanAttentionBlock):
+    def __init__(
+            self,
+            cross_attn_type,
+            dim,
+            ffn_dim,
+            num_heads,
+            window_size=(-1, -1),
+            qk_norm=True,
+            cross_attn_norm=False,
+            eps=1e-6,
+            block_id=0,
+            operation_settings={}
+    ):
+        super().__init__(cross_attn_type, dim, ffn_dim, num_heads, window_size, qk_norm, cross_attn_norm, eps, operation_settings=operation_settings)
+        self.block_id = block_id
+        if block_id == 0:
+            self.before_proj = operation_settings.get("operations").Linear(self.dim, self.dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.after_proj = operation_settings.get("operations").Linear(self.dim, self.dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, c, x, **kwargs):
+        if self.block_id == 0:
+            c = self.before_proj(c) + x
+        c = super().forward(c, **kwargs)
+        c_skip = self.after_proj(c)
+        return c_skip, c
+
+
+class WanCamAdapter(nn.Module):
+    def __init__(self, in_dim, out_dim, kernel_size, stride, num_residual_blocks=1, operation_settings={}):
+        super(WanCamAdapter, self).__init__()
+
+        # Pixel Unshuffle: reduce spatial dimensions by a factor of 8
+        self.pixel_unshuffle = nn.PixelUnshuffle(downscale_factor=8)
+
+        # Convolution: reduce spatial dimensions by a factor
+        #  of 2 (without overlap)
+        self.conv = operation_settings.get("operations").Conv2d(in_dim * 64, out_dim, kernel_size=kernel_size, stride=stride, padding=0, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+        # Residual blocks for feature extraction
+        self.residual_blocks = nn.Sequential(
+            *[WanCamResidualBlock(out_dim, operation_settings = operation_settings) for _ in range(num_residual_blocks)]
+        )
+
+    def forward(self, x):
+        # Reshape to merge the frame dimension into batch
+        bs, c, f, h, w = x.size()
+        x = x.permute(0, 2, 1, 3, 4).contiguous().view(bs * f, c, h, w)
+
+        # Pixel Unshuffle operation
+        x_unshuffled = self.pixel_unshuffle(x)
+
+        # Convolution operation
+        x_conv = self.conv(x_unshuffled)
+
+        # Feature extraction with residual blocks
+        out = self.residual_blocks(x_conv)
+
+        # Reshape to restore original bf dimension
+        out = out.view(bs, f, out.size(1), out.size(2), out.size(3))
+
+        # Permute dimensions to reorder (if needed), e.g., swap channels and feature frames
+        out = out.permute(0, 2, 1, 3, 4)
+
+        return out
+
+
+class WanCamResidualBlock(nn.Module):
+    def __init__(self, dim, operation_settings={}):
+        super(WanCamResidualBlock, self).__init__()
+        self.conv1 = operation_settings.get("operations").Conv2d(dim, dim, kernel_size=3, padding=1, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.relu = nn.ReLU(inplace=True)
+        self.conv2 = operation_settings.get("operations").Conv2d(dim, dim, kernel_size=3, padding=1, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+    def forward(self, x):
+        residual = x
+        out = self.relu(self.conv1(x))
+        out = self.conv2(out)
+        out += residual
+        return out
+
+
+class Head(nn.Module):
+
+    def __init__(self, dim, out_dim, patch_size, eps=1e-6, operation_settings={}):
+        super().__init__()
+        self.dim = dim
+        self.out_dim = out_dim
+        self.patch_size = patch_size
+        self.eps = eps
+
+        # layers
+        out_dim = math.prod(patch_size) * out_dim
+        self.norm = operation_settings.get("operations").LayerNorm(dim, eps, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        # self.norm = LayerNorm(dim, eps, elementwise_affine=False, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        self.head = operation_settings.get("operations").Linear(dim, out_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+
+        # modulation
+        self.modulation = nn.Parameter(torch.empty(1, 2, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+    def forward(self, x, e):
+        r"""
+        Args:
+            x(Tensor): Shape [B, L1, C]
+            e(Tensor): Shape [B, C]
+        """
+        # assert e.dtype == torch.float32
+        if e.ndim < 3:
+            e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device) + e.unsqueeze(1)).chunk(2, dim=1)
+        else:
+            e = (comfy.model_management.cast_to(self.modulation, dtype=x.dtype, device=x.device).unsqueeze(0) + e.unsqueeze(2)).unbind(2)
+
+        x = (self.head(torch.addcmul(repeat_e(e[0], x), self.norm(x), 1 + repeat_e(e[1], x))))
+        return x
+
+
+class MLPProj(torch.nn.Module):
+
+    def __init__(self, in_dim, out_dim, flf_pos_embed_token_number=None, operation_settings={}):
+        super().__init__()
+
+        self.proj = torch.nn.Sequential(
+            operation_settings.get("operations").LayerNorm(in_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), operation_settings.get("operations").Linear(in_dim, in_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
+            torch.nn.GELU(), operation_settings.get("operations").Linear(in_dim, out_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")),
+            operation_settings.get("operations").LayerNorm(out_dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+        if flf_pos_embed_token_number is not None:
+            self.emb_pos = nn.Parameter(torch.empty((1, flf_pos_embed_token_number, in_dim), device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+        else:
+            self.emb_pos = None
+
+    def forward(self, image_embeds):
+        if self.emb_pos is not None:
+            image_embeds = image_embeds[:, :self.emb_pos.shape[1]] + comfy.model_management.cast_to(self.emb_pos[:, :image_embeds.shape[1]], dtype=image_embeds.dtype, device=image_embeds.device)
+
+        clip_extra_context_tokens = self.proj(image_embeds)
+        return clip_extra_context_tokens
+
+
+class WanModel(torch.nn.Module):
+    r"""
+    Wan diffusion backbone supporting both text-to-video and image-to-video.
+    """
+
+    def __init__(self,
+                 model_type='t2v',
+                 patch_size=(1, 2, 2),
+                 text_len=512,
+                 in_dim=16,
+                 dim=2048,
+                 ffn_dim=8192,
+                 freq_dim=256,
+                 text_dim=4096,
+                 out_dim=16,
+                 num_heads=16,
+                 num_layers=32,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=True,
+                 eps=1e-6,
+                 flf_pos_embed_token_number=None,
+                 in_dim_ref_conv=None,
+                 image_model=None,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+        r"""
+        Initialize the diffusion model backbone.
+
+        Args:
+            model_type (`str`, *optional*, defaults to 't2v'):
+                Model variant - 't2v' (text-to-video) or 'i2v' (image-to-video)
+            patch_size (`tuple`, *optional*, defaults to (1, 2, 2)):
+                3D patch dimensions for video embedding (t_patch, h_patch, w_patch)
+            text_len (`int`, *optional*, defaults to 512):
+                Fixed length for text embeddings
+            in_dim (`int`, *optional*, defaults to 16):
+                Input video channels (C_in)
+            dim (`int`, *optional*, defaults to 2048):
+                Hidden dimension of the transformer
+            ffn_dim (`int`, *optional*, defaults to 8192):
+                Intermediate dimension in feed-forward network
+            freq_dim (`int`, *optional*, defaults to 256):
+                Dimension for sinusoidal time embeddings
+            text_dim (`int`, *optional*, defaults to 4096):
+                Input dimension for text embeddings
+            out_dim (`int`, *optional*, defaults to 16):
+                Output video channels (C_out)
+            num_heads (`int`, *optional*, defaults to 16):
+                Number of attention heads
+            num_layers (`int`, *optional*, defaults to 32):
+                Number of transformer blocks
+            window_size (`tuple`, *optional*, defaults to (-1, -1)):
+                Window size for local attention (-1 indicates global attention)
+            qk_norm (`bool`, *optional*, defaults to True):
+                Enable query/key normalization
+            cross_attn_norm (`bool`, *optional*, defaults to False):
+                Enable cross-attention normalization
+            eps (`float`, *optional*, defaults to 1e-6):
+                Epsilon value for normalization layers
+        """
+
+        super().__init__()
+        self.dtype = dtype
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+
+        assert model_type in ['t2v', 'i2v']
+        self.model_type = model_type
+
+        self.patch_size = patch_size
+        self.text_len = text_len
+        self.in_dim = in_dim
+        self.dim = dim
+        self.ffn_dim = ffn_dim
+        self.freq_dim = freq_dim
+        self.text_dim = text_dim
+        self.out_dim = out_dim
+        self.num_heads = num_heads
+        self.num_layers = num_layers
+        self.window_size = window_size
+        self.qk_norm = qk_norm
+        self.cross_attn_norm = cross_attn_norm
+        self.eps = eps
+
+        # embeddings
+        self.patch_embedding = operations.Conv3d(
+            in_dim, dim, kernel_size=patch_size, stride=patch_size, device=operation_settings.get("device"), dtype=torch.float32)
+        self.text_embedding = nn.Sequential(
+            operations.Linear(text_dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), nn.GELU(approximate='tanh'),
+            operations.Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+        self.time_embedding = nn.Sequential(
+            operations.Linear(freq_dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")), nn.SiLU(), operations.Linear(dim, dim, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+        self.time_projection = nn.Sequential(nn.SiLU(), operations.Linear(dim, dim * 6, device=operation_settings.get("device"), dtype=operation_settings.get("dtype")))
+
+        # blocks
+        cross_attn_type = 't2v_cross_attn' if model_type == 't2v' else 'i2v_cross_attn'
+        self.blocks = nn.ModuleList([
+            WanAttentionBlock(cross_attn_type, dim, ffn_dim, num_heads,
+                              window_size, qk_norm, cross_attn_norm, eps, operation_settings=operation_settings)
+            for _ in range(num_layers)
+        ])
+
+        # head
+        self.head = Head(dim, out_dim, patch_size, eps, operation_settings=operation_settings)
+
+        d = dim // num_heads
+        self.rope_embedder = EmbedND(dim=d, theta=10000.0, axes_dim=[d - 4 * (d // 6), 2 * (d // 6), 2 * (d // 6)])
+
+        if model_type == 'i2v':
+            self.img_emb = MLPProj(1280, dim, flf_pos_embed_token_number=flf_pos_embed_token_number, operation_settings=operation_settings)
+        else:
+            self.img_emb = None
+
+        if in_dim_ref_conv is not None:
+            self.ref_conv = operations.Conv2d(in_dim_ref_conv, dim, kernel_size=patch_size[1:], stride=patch_size[1:], device=operation_settings.get("device"), dtype=operation_settings.get("dtype"))
+        else:
+            self.ref_conv = None
+
+    def forward_orig(
+        self,
+        x,
+        t,
+        context,
+        clip_fea=None,
+        freqs=None,
+        transformer_options={},
+        **kwargs,
+    ):
+        r"""
+        Forward pass through the diffusion model
+
+        Args:
+            x (Tensor):
+                List of input video tensors with shape [B, C_in, F, H, W]
+            t (Tensor):
+                Diffusion timesteps tensor of shape [B]
+            context (List[Tensor]):
+                List of text embeddings each with shape [B, L, C]
+            seq_len (`int`):
+                Maximum sequence length for positional encoding
+            clip_fea (Tensor, *optional*):
+                CLIP image features for image-to-video mode
+            y (List[Tensor], *optional*):
+                Conditional video inputs for image-to-video mode, same shape as x
+
+        Returns:
+            List[Tensor]:
+                List of denoised video tensors with original input shapes [C_out, F, H / 8, W / 8]
+        """
+        # embeddings
+        # import pdb
+        # pdb.set_trace()
+        x = self.patch_embedding(x.float()).to(x.dtype)
+        grid_sizes = x.shape[2:]
+        x = x.flatten(2).transpose(1, 2).contiguous()   # chenxi add .contiguous()
+
+        # time embeddings
+        e = self.time_embedding(
+            sinusoidal_embedding_1d(self.freq_dim, t.flatten()).to(dtype=x[0].dtype))
+       
+        e = e.reshape(t.shape[0], -1, e.shape[-1])
+        e0 = self.time_projection(e).unflatten(2, (6, self.dim))
+
+        full_ref = None
+        if self.ref_conv is not None:
+            full_ref = kwargs.get("reference_latent", None)
+            if full_ref is not None:
+                full_ref = self.ref_conv(full_ref).flatten(2).transpose(1, 2)
+                x = torch.concat((full_ref, x), dim=1)
+
+        # context
+        context = self.text_embedding(context)
+
+        context_img_len = None
+        if clip_fea is not None:
+            if self.img_emb is not None:
+                context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
+                context = torch.concat([context_clip, context], dim=1)
+            context_img_len = clip_fea.shape[-2]
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len)
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
+
+        # head
+        x = self.head(x, e)
+
+        if full_ref is not None:
+            x = x[:, full_ref.shape[1]:]
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return x
+
+    def rope_encode(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None):
+        patch_size = self.patch_size
+        t_len = ((t + (patch_size[0] // 2)) // patch_size[0])
+        h_len = ((h + (patch_size[1] // 2)) // patch_size[1])
+        w_len = ((w + (patch_size[2] // 2)) // patch_size[2])
+
+        if steps_t is None:
+            steps_t = t_len
+        if steps_h is None:
+            steps_h = h_len
+        if steps_w is None:
+            steps_w = w_len
+
+        img_ids = torch.zeros((steps_t, steps_h, steps_w, 3), device=device, dtype=dtype)
+        img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(t_start, t_start + (t_len - 1), steps=steps_t, device=device, dtype=dtype).reshape(-1, 1, 1)
+        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(0, h_len - 1, steps=steps_h, device=device, dtype=dtype).reshape(1, -1, 1)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(0, w_len - 1, steps=steps_w, device=device, dtype=dtype).reshape(1, 1, -1)
+        img_ids = img_ids.reshape(1, -1, img_ids.shape[-1])
+
+        freqs = self.rope_embedder(img_ids).movedim(1, 2)
+        return freqs
+
+    def forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._forward,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
+        ).execute(x, timestep, context, clip_fea, time_dim_concat, transformer_options, **kwargs)
+
+    def _forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, **kwargs):
+        bs, c, t, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
+
+        t_len = t
+        if time_dim_concat is not None:
+            time_dim_concat = comfy.ldm.common_dit.pad_to_patch_size(time_dim_concat, self.patch_size)
+            x = torch.cat([x, time_dim_concat], dim=2)
+            t_len = x.shape[2]
+
+        if self.ref_conv is not None and "reference_latent" in kwargs:
+            t_len += 1
+
+        freqs = self.rope_encode(t_len, h, w, device=x.device, dtype=x.dtype)
+        return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs, transformer_options=transformer_options, **kwargs)[:, :, :t, :h, :w]
+
+    def unpatchify(self, x, grid_sizes):
+        r"""
+        Reconstruct video tensors from patch embeddings.
+
+        Args:
+            x (List[Tensor]):
+                List of patchified features, each with shape [L, C_out * prod(patch_size)]
+            grid_sizes (Tensor):
+                Original spatial-temporal grid dimensions before patching,
+                    shape [B, 3] (3 dimensions correspond to F_patches, H_patches, W_patches)
+
+        Returns:
+            List[Tensor]:
+                Reconstructed video tensors with shape [L, C_out, F, H / 8, W / 8]
+        """
+
+        c = self.out_dim
+        u = x
+        b = u.shape[0]
+        u = u[:, :math.prod(grid_sizes)].view(b, *grid_sizes, *self.patch_size, c)
+        u = torch.einsum('bfhwpqrc->bcfphqwr', u)
+        u = u.reshape(b, c, *[i * j for i, j in zip(grid_sizes, self.patch_size)])
+        return u
+
+
+class VaceWanModel(WanModel):
+    r"""
+    Wan diffusion backbone supporting both text-to-video and image-to-video.
+    """
+
+    def __init__(self,
+                 model_type='vace',
+                 patch_size=(1, 2, 2),
+                 text_len=512,
+                 in_dim=16,
+                 dim=2048,
+                 ffn_dim=8192,
+                 freq_dim=256,
+                 text_dim=4096,
+                 out_dim=16,
+                 num_heads=16,
+                 num_layers=32,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=True,
+                 eps=1e-6,
+                 flf_pos_embed_token_number=None,
+                 image_model=None,
+                 vace_layers=None,
+                 vace_in_dim=None,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+
+        super().__init__(model_type='t2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations)
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+
+        # Vace
+        if vace_layers is not None:
+            self.vace_layers = vace_layers
+            self.vace_in_dim = vace_in_dim
+            # vace blocks
+            self.vace_blocks = nn.ModuleList([
+                VaceWanAttentionBlock('t2v_cross_attn', self.dim, self.ffn_dim, self.num_heads, self.window_size, self.qk_norm, self.cross_attn_norm, self.eps, block_id=i, operation_settings=operation_settings)
+                for i in range(self.vace_layers)
+            ])
+
+            self.vace_layers_mapping = {i: n for n, i in enumerate(range(0, self.num_layers, self.num_layers // self.vace_layers))}
+            # vace patch embeddings
+            self.vace_patch_embedding = operations.Conv3d(
+                self.vace_in_dim, self.dim, kernel_size=self.patch_size, stride=self.patch_size, device=device, dtype=torch.float32
+            )
+
+    def forward_orig(
+        self,
+        x,
+        t,
+        context,
+        vace_context,
+        vace_strength,
+        clip_fea=None,
+        freqs=None,
+        transformer_options={},
+        **kwargs,
+    ):
+        # embeddings
+        x = self.patch_embedding(x.float()).to(x.dtype)
+        grid_sizes = x.shape[2:]
+        x = x.flatten(2).transpose(1, 2)
+
+        # time embeddings
+        e = self.time_embedding(
+            sinusoidal_embedding_1d(self.freq_dim, t).to(dtype=x[0].dtype))
+        e0 = self.time_projection(e).unflatten(1, (6, self.dim))
+
+        # context
+        context = self.text_embedding(context)
+
+        context_img_len = None
+        if clip_fea is not None:
+            if self.img_emb is not None:
+                context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
+                context = torch.concat([context_clip, context], dim=1)
+            context_img_len = clip_fea.shape[-2]
+
+        orig_shape = list(vace_context.shape)
+        vace_context = vace_context.movedim(0, 1).reshape([-1] + orig_shape[2:])
+        c = self.vace_patch_embedding(vace_context.float()).to(vace_context.dtype)
+        c = c.flatten(2).transpose(1, 2)
+        c = list(c.split(orig_shape[0], dim=0))
+
+        # arguments
+        x_orig = x
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len)
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
+
+            ii = self.vace_layers_mapping.get(i, None)
+            if ii is not None:
+                for iii in range(len(c)):
+                    c_skip, c[iii] = self.vace_blocks[ii](c[iii], x=x_orig, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
+                    x += c_skip * vace_strength[iii]
+                del c_skip
+        # head
+        x = self.head(x, e)
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return x
+
+class CameraWanModel(WanModel):
+    r"""
+    Wan diffusion backbone supporting both text-to-video and image-to-video.
+    """
+
+    def __init__(self,
+                 model_type='camera',
+                 patch_size=(1, 2, 2),
+                 text_len=512,
+                 in_dim=16,
+                 dim=2048,
+                 ffn_dim=8192,
+                 freq_dim=256,
+                 text_dim=4096,
+                 out_dim=16,
+                 num_heads=16,
+                 num_layers=32,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=True,
+                 eps=1e-6,
+                 flf_pos_embed_token_number=None,
+                 image_model=None,
+                 in_dim_control_adapter=24,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+
+        if model_type == 'camera':
+            model_type = 'i2v'
+        else:
+            model_type = 't2v'
+
+        super().__init__(model_type=model_type, patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, flf_pos_embed_token_number=flf_pos_embed_token_number, image_model=image_model, device=device, dtype=dtype, operations=operations)
+        operation_settings = {"operations": operations, "device": device, "dtype": dtype}
+
+        self.control_adapter = WanCamAdapter(in_dim_control_adapter, dim, kernel_size=patch_size[1:], stride=patch_size[1:], operation_settings=operation_settings)
+
+
+    def forward_orig(
+        self,
+        x,
+        t,
+        context,
+        clip_fea=None,
+        freqs=None,
+        camera_conditions = None,
+        transformer_options={},
+        **kwargs,
+    ):
+        # embeddings
+        x = self.patch_embedding(x.float()).to(x.dtype)
+        if self.control_adapter is not None and camera_conditions is not None:
+            x = x + self.control_adapter(camera_conditions).to(x.dtype)
+        grid_sizes = x.shape[2:]
+        x = x.flatten(2).transpose(1, 2)
+
+        # time embeddings
+        e = self.time_embedding(
+            sinusoidal_embedding_1d(self.freq_dim, t).to(dtype=x[0].dtype))
+        e0 = self.time_projection(e).unflatten(1, (6, self.dim))
+
+        # context
+        context = self.text_embedding(context)
+
+        context_img_len = None
+        if clip_fea is not None:
+            if self.img_emb is not None:
+                context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
+                context = torch.concat([context_clip, context], dim=1)
+            context_img_len = clip_fea.shape[-2]
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len)
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len)
+
+        # head
+        x = self.head(x, e)
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return x
+
+
+class CausalConv1d(nn.Module):
+
+    def __init__(self,
+                 chan_in,
+                 chan_out,
+                 kernel_size=3,
+                 stride=1,
+                 dilation=1,
+                 pad_mode='replicate',
+                 operations=None,
+                 **kwargs):
+        super().__init__()
+
+        self.pad_mode = pad_mode
+        padding = (kernel_size - 1, 0)  # T
+        self.time_causal_padding = padding
+
+        self.conv = operations.Conv1d(
+            chan_in,
+            chan_out,
+            kernel_size,
+            stride=stride,
+            dilation=dilation,
+            **kwargs)
+
+    def forward(self, x):
+        x = torch.nn.functional.pad(x, self.time_causal_padding, mode=self.pad_mode)
+        return self.conv(x)
+
+
+class MotionEncoder_tc(nn.Module):
+
+    def __init__(self,
+                 in_dim: int,
+                 hidden_dim: int,
+                 num_heads=int,
+                 need_global=True,
+                 dtype=None,
+                 device=None,
+                 operations=None,):
+        factory_kwargs = {"dtype": dtype, "device": device}
+        super().__init__()
+
+        self.num_heads = num_heads
+        self.need_global = need_global
+        self.conv1_local = CausalConv1d(in_dim, hidden_dim // 4 * num_heads, 3, stride=1, operations=operations, **factory_kwargs)
+        if need_global:
+            self.conv1_global = CausalConv1d(
+                in_dim, hidden_dim // 4, 3, stride=1, operations=operations, **factory_kwargs)
+        self.norm1 = operations.LayerNorm(
+            hidden_dim // 4,
+            elementwise_affine=False,
+            eps=1e-6,
+            **factory_kwargs)
+        self.act = nn.SiLU()
+        self.conv2 = CausalConv1d(hidden_dim // 4, hidden_dim // 2, 3, stride=2, operations=operations, **factory_kwargs)
+        self.conv3 = CausalConv1d(hidden_dim // 2, hidden_dim, 3, stride=2, operations=operations, **factory_kwargs)
+
+        if need_global:
+            self.final_linear = operations.Linear(hidden_dim, hidden_dim, **factory_kwargs)
+
+        self.norm1 = operations.LayerNorm(
+            hidden_dim // 4,
+            elementwise_affine=False,
+            eps=1e-6,
+            **factory_kwargs)
+
+        self.norm2 = operations.LayerNorm(
+            hidden_dim // 2,
+            elementwise_affine=False,
+            eps=1e-6,
+            **factory_kwargs)
+
+        self.norm3 = operations.LayerNorm(
+            hidden_dim, elementwise_affine=False, eps=1e-6, **factory_kwargs)
+
+        self.padding_tokens = nn.Parameter(torch.empty(1, 1, 1, hidden_dim, **factory_kwargs))
+
+    def forward(self, x):
+        x = rearrange(x, 'b t c -> b c t')
+        x_ori = x.clone()
+        b, c, t = x.shape
+        x = self.conv1_local(x)
+        x = rearrange(x, 'b (n c) t -> (b n) t c', n=self.num_heads)
+        x = self.norm1(x)
+        x = self.act(x)
+        x = rearrange(x, 'b t c -> b c t')
+        x = self.conv2(x)
+        x = rearrange(x, 'b c t -> b t c')
+        x = self.norm2(x)
+        x = self.act(x)
+        x = rearrange(x, 'b t c -> b c t')
+        x = self.conv3(x)
+        x = rearrange(x, 'b c t -> b t c')
+        x = self.norm3(x)
+        x = self.act(x)
+        x = rearrange(x, '(b n) t c -> b t n c', b=b)
+        padding = comfy.model_management.cast_to(self.padding_tokens, dtype=x.dtype, device=x.device).repeat(b, x.shape[1], 1, 1)
+        x = torch.cat([x, padding], dim=-2)
+        x_local = x.clone()
+
+        if not self.need_global:
+            return x_local
+
+        x = self.conv1_global(x_ori)
+        x = rearrange(x, 'b c t -> b t c')
+        x = self.norm1(x)
+        x = self.act(x)
+        x = rearrange(x, 'b t c -> b c t')
+        x = self.conv2(x)
+        x = rearrange(x, 'b c t -> b t c')
+        x = self.norm2(x)
+        x = self.act(x)
+        x = rearrange(x, 'b t c -> b c t')
+        x = self.conv3(x)
+        x = rearrange(x, 'b c t -> b t c')
+        x = self.norm3(x)
+        x = self.act(x)
+        x = self.final_linear(x)
+        x = rearrange(x, '(b n) t c -> b t n c', b=b)
+
+        return x, x_local
+
+
+class CausalAudioEncoder(nn.Module):
+
+    def __init__(self,
+                 dim=5120,
+                 num_layers=25,
+                 out_dim=2048,
+                 video_rate=8,
+                 num_token=4,
+                 need_global=False,
+                 dtype=None,
+                 device=None,
+                 operations=None):
+        super().__init__()
+        self.encoder = MotionEncoder_tc(
+            in_dim=dim,
+            hidden_dim=out_dim,
+            num_heads=num_token,
+            need_global=need_global, dtype=dtype, device=device, operations=operations)
+        weight = torch.empty((1, num_layers, 1, 1), dtype=dtype, device=device)
+
+        self.weights = torch.nn.Parameter(weight)
+        self.act = torch.nn.SiLU()
+
+    def forward(self, features):
+        # features B * num_layers * dim * video_length
+        weights = self.act(comfy.model_management.cast_to(self.weights, dtype=features.dtype, device=features.device))
+        weights_sum = weights.sum(dim=1, keepdims=True)
+        weighted_feat = ((features * weights) / weights_sum).sum(
+            dim=1)  # b dim f
+        weighted_feat = weighted_feat.permute(0, 2, 1)  # b f dim
+        res = self.encoder(weighted_feat)  # b f n dim
+        return res  # b f n dim
+
+
+class AdaLayerNorm(nn.Module):
+    def __init__(self, embedding_dim, output_dim=None, norm_elementwise_affine=False, norm_eps=1e-5, dtype=None, device=None, operations=None):
+        super().__init__()
+
+        output_dim = output_dim or embedding_dim * 2
+
+        self.silu = nn.SiLU()
+        self.linear = operations.Linear(embedding_dim, output_dim, dtype=dtype, device=device)
+        self.norm = operations.LayerNorm(output_dim // 2, norm_eps, norm_elementwise_affine, dtype=dtype, device=device)
+
+    def forward(self, x, temb):
+        temb = self.linear(self.silu(temb))
+        shift, scale = temb.chunk(2, dim=1)
+        shift = shift[:, None, :]
+        scale = scale[:, None, :]
+        x = self.norm(x) * (1 + scale) + shift
+        return x
+
+
+class AudioInjector_WAN(nn.Module):
+
+    def __init__(self,
+                 dim=2048,
+                 num_heads=32,
+                 inject_layer=[0, 27],
+                 root_net=None,
+                 enable_adain=False,
+                 adain_dim=2048,
+                 adain_mode=None,
+                 dtype=None,
+                 device=None,
+                 operations=None):
+        super().__init__()
+        self.enable_adain = enable_adain
+        self.adain_mode = adain_mode
+        self.injected_block_id = {}
+        audio_injector_id = 0
+        for inject_id in inject_layer:
+            self.injected_block_id[inject_id] = audio_injector_id
+            audio_injector_id += 1
+
+        self.injector = nn.ModuleList([
+            WanT2VCrossAttention(
+                dim=dim,
+                num_heads=num_heads,
+                qk_norm=True, operation_settings={"operations": operations, "device": device, "dtype": dtype}
+            ) for _ in range(audio_injector_id)
+        ])
+        self.injector_pre_norm_feat = nn.ModuleList([
+            operations.LayerNorm(
+                dim,
+                elementwise_affine=False,
+                eps=1e-6, dtype=dtype, device=device
+            ) for _ in range(audio_injector_id)
+        ])
+        self.injector_pre_norm_vec = nn.ModuleList([
+            operations.LayerNorm(
+                dim,
+                elementwise_affine=False,
+                eps=1e-6, dtype=dtype, device=device
+            ) for _ in range(audio_injector_id)
+        ])
+        if enable_adain:
+            self.injector_adain_layers = nn.ModuleList([
+                AdaLayerNorm(
+                    output_dim=dim * 2, embedding_dim=adain_dim, dtype=dtype, device=device, operations=operations)
+                for _ in range(audio_injector_id)
+            ])
+            if adain_mode != "attn_norm":
+                self.injector_adain_output_layers = nn.ModuleList(
+                    [operations.Linear(dim, dim, dtype=dtype, device=device) for _ in range(audio_injector_id)])
+
+    def forward(self, x, block_id, audio_emb, audio_emb_global, seq_len):
+        audio_attn_id = self.injected_block_id.get(block_id, None)
+        if audio_attn_id is None:
+            return x
+
+        num_frames = audio_emb.shape[1]
+        input_hidden_states = rearrange(x[:, :seq_len], "b (t n) c -> (b t) n c", t=num_frames)
+        if self.enable_adain and self.adain_mode == "attn_norm":
+            audio_emb_global = rearrange(audio_emb_global, "b t n c -> (b t) n c")
+            adain_hidden_states = self.injector_adain_layers[audio_attn_id](input_hidden_states, temb=audio_emb_global[:, 0])
+            attn_hidden_states = adain_hidden_states
+        else:
+            attn_hidden_states = self.injector_pre_norm_feat[audio_attn_id](input_hidden_states)
+        audio_emb = rearrange(audio_emb, "b t n c -> (b t) n c", t=num_frames)
+        attn_audio_emb = audio_emb
+        residual_out = self.injector[audio_attn_id](x=attn_hidden_states, context=attn_audio_emb)
+        residual_out = rearrange(
+            residual_out, "(b t) n c -> b (t n) c", t=num_frames)
+        x[:, :seq_len] = x[:, :seq_len] + residual_out
+        return x
+
+
+class FramePackMotioner(nn.Module):
+    def __init__(
+            self,
+            inner_dim=1024,
+            num_heads=16,  # Used to indicate the number of heads in the backbone network; unrelated to this module's design
+            zip_frame_buckets=[
+                1, 2, 16
+            ],  # Three numbers representing the number of frames sampled for patch operations from the nearest to the farthest frames
+            drop_mode="drop",  # If not "drop", it will use "padd", meaning padding instead of deletion
+            dtype=None,
+            device=None,
+            operations=None):
+        super().__init__()
+        self.proj = operations.Conv3d(16, inner_dim, kernel_size=(1, 2, 2), stride=(1, 2, 2), dtype=dtype, device=device)
+        self.proj_2x = operations.Conv3d(16, inner_dim, kernel_size=(2, 4, 4), stride=(2, 4, 4), dtype=dtype, device=device)
+        self.proj_4x = operations.Conv3d(16, inner_dim, kernel_size=(4, 8, 8), stride=(4, 8, 8), dtype=dtype, device=device)
+        self.zip_frame_buckets = zip_frame_buckets
+
+        self.inner_dim = inner_dim
+        self.num_heads = num_heads
+
+        self.drop_mode = drop_mode
+
+    def forward(self, motion_latents, rope_embedder, add_last_motion=2):
+        lat_height, lat_width = motion_latents.shape[3], motion_latents.shape[4]
+        padd_lat = torch.zeros(motion_latents.shape[0], 16, sum(self.zip_frame_buckets), lat_height, lat_width).to(device=motion_latents.device, dtype=motion_latents.dtype)
+        overlap_frame = min(padd_lat.shape[2], motion_latents.shape[2])
+        if overlap_frame > 0:
+            padd_lat[:, :, -overlap_frame:] = motion_latents[:, :, -overlap_frame:]
+
+        if add_last_motion < 2 and self.drop_mode != "drop":
+            zero_end_frame = sum(self.zip_frame_buckets[:len(self.zip_frame_buckets) - add_last_motion - 1])
+            padd_lat[:, :, -zero_end_frame:] = 0
+
+        clean_latents_4x, clean_latents_2x, clean_latents_post = padd_lat[:, :, -sum(self.zip_frame_buckets):, :, :].split(self.zip_frame_buckets[::-1], dim=2)  # 16, 2 ,1
+
+        # patchfy
+        clean_latents_post = self.proj(clean_latents_post).flatten(2).transpose(1, 2)
+        clean_latents_2x = self.proj_2x(clean_latents_2x)
+        l_2x_shape = clean_latents_2x.shape
+        clean_latents_2x = clean_latents_2x.flatten(2).transpose(1, 2)
+        clean_latents_4x = self.proj_4x(clean_latents_4x)
+        l_4x_shape = clean_latents_4x.shape
+        clean_latents_4x = clean_latents_4x.flatten(2).transpose(1, 2)
+
+        if add_last_motion < 2 and self.drop_mode == "drop":
+            clean_latents_post = clean_latents_post[:, :
+                                                    0] if add_last_motion < 2 else clean_latents_post
+            clean_latents_2x = clean_latents_2x[:, :
+                                                0] if add_last_motion < 1 else clean_latents_2x
+
+        motion_lat = torch.cat([clean_latents_post, clean_latents_2x, clean_latents_4x], dim=1)
+
+        rope_post = rope_embedder.rope_encode(1, lat_height, lat_width, t_start=-1, device=motion_latents.device, dtype=motion_latents.dtype)
+        rope_2x = rope_embedder.rope_encode(1, lat_height, lat_width, t_start=-3, steps_h=l_2x_shape[-2], steps_w=l_2x_shape[-1], device=motion_latents.device, dtype=motion_latents.dtype)
+        rope_4x = rope_embedder.rope_encode(4, lat_height, lat_width, t_start=-19, steps_h=l_4x_shape[-2], steps_w=l_4x_shape[-1], device=motion_latents.device, dtype=motion_latents.dtype)
+
+        rope = torch.cat([rope_post, rope_2x, rope_4x], dim=1)
+        return motion_lat, rope
+
+
+class WanModel_S2V(WanModel):
+    def __init__(self,
+                 model_type='s2v',
+                 patch_size=(1, 2, 2),
+                 text_len=512,
+                 in_dim=16,
+                 dim=2048,
+                 ffn_dim=8192,
+                 freq_dim=256,
+                 text_dim=4096,
+                 out_dim=16,
+                 num_heads=16,
+                 num_layers=32,
+                 window_size=(-1, -1),
+                 qk_norm=True,
+                 cross_attn_norm=True,
+                 eps=1e-6,
+                 audio_dim=1024,
+                 num_audio_token=4,
+                 enable_adain=True,
+                 cond_dim=16,
+                 audio_inject_layers=[0, 4, 8, 12, 16, 20, 24, 27, 30, 33, 36, 39],
+                 adain_mode="attn_norm",
+                 framepack_drop_mode="padd",
+                 image_model=None,
+                 device=None,
+                 dtype=None,
+                 operations=None,
+                 ):
+
+        super().__init__(model_type='t2v', patch_size=patch_size, text_len=text_len, in_dim=in_dim, dim=dim, ffn_dim=ffn_dim, freq_dim=freq_dim, text_dim=text_dim, out_dim=out_dim, num_heads=num_heads, num_layers=num_layers, window_size=window_size, qk_norm=qk_norm, cross_attn_norm=cross_attn_norm, eps=eps, image_model=image_model, device=device, dtype=dtype, operations=operations)
+
+        self.trainable_cond_mask = operations.Embedding(3, self.dim, device=device, dtype=dtype)
+
+        self.casual_audio_encoder = CausalAudioEncoder(
+            dim=audio_dim,
+            out_dim=self.dim,
+            num_token=num_audio_token,
+            need_global=enable_adain, dtype=dtype, device=device, operations=operations)
+
+        if cond_dim > 0:
+            self.cond_encoder = operations.Conv3d(
+                cond_dim,
+                self.dim,
+                kernel_size=self.patch_size,
+                stride=self.patch_size, device=device, dtype=dtype)
+
+        self.audio_injector = AudioInjector_WAN(
+            dim=self.dim,
+            num_heads=self.num_heads,
+            inject_layer=audio_inject_layers,
+            root_net=self,
+            enable_adain=enable_adain,
+            adain_dim=self.dim,
+            adain_mode=adain_mode,
+            dtype=dtype, device=device, operations=operations
+        )
+
+        self.frame_packer = FramePackMotioner(
+            inner_dim=self.dim,
+            num_heads=self.num_heads,
+            zip_frame_buckets=[1, 2, 16],
+            drop_mode=framepack_drop_mode,
+            dtype=dtype, device=device, operations=operations)
+
+    def forward_orig(
+        self,
+        x,
+        t,
+        context,
+        audio_embed=None,
+        reference_latent=None,
+        control_video=None,
+        reference_motion=None,
+        clip_fea=None,
+        freqs=None,
+        transformer_options={},
+        **kwargs,
+    ):
+        if audio_embed is not None:
+            num_embeds = x.shape[-3] * 4
+            audio_emb_global, audio_emb = self.casual_audio_encoder(audio_embed[:, :, :, :num_embeds])
+        else:
+            audio_emb = None
+
+        # embeddings
+        bs, _, time, height, width = x.shape
+        x = self.patch_embedding(x.float()).to(x.dtype)
+        if control_video is not None:
+            x = x + self.cond_encoder(control_video)
+
+        if t.ndim == 1:
+            t = t.unsqueeze(1).repeat(1, x.shape[2])
+
+        grid_sizes = x.shape[2:]
+        x = x.flatten(2).transpose(1, 2)
+        seq_len = x.size(1)
+
+        cond_mask_weight = comfy.model_management.cast_to(self.trainable_cond_mask.weight, dtype=x.dtype, device=x.device).unsqueeze(1).unsqueeze(1)
+        x = x + cond_mask_weight[0]
+
+        if reference_latent is not None:
+            ref = self.patch_embedding(reference_latent.float()).to(x.dtype)
+            ref = ref.flatten(2).transpose(1, 2)
+            freqs_ref = self.rope_encode(reference_latent.shape[-3], reference_latent.shape[-2], reference_latent.shape[-1], t_start=max(30, time + 9), device=x.device, dtype=x.dtype)
+            ref = ref + cond_mask_weight[1]
+            x = torch.cat([x, ref], dim=1)
+            freqs = torch.cat([freqs, freqs_ref], dim=1)
+            t = torch.cat([t, torch.zeros((t.shape[0], reference_latent.shape[-3]), device=t.device, dtype=t.dtype)], dim=1)
+            del ref, freqs_ref
+
+        if reference_motion is not None:
+            motion_encoded, freqs_motion = self.frame_packer(reference_motion, self)
+            motion_encoded = motion_encoded + cond_mask_weight[2]
+            x = torch.cat([x, motion_encoded], dim=1)
+            freqs = torch.cat([freqs, freqs_motion], dim=1)
+
+            t = torch.repeat_interleave(t, 2, dim=1)
+            t = torch.cat([t, torch.zeros((t.shape[0], 3), device=t.device, dtype=t.dtype)], dim=1)
+            del motion_encoded, freqs_motion
+
+        # time embeddings
+        e = self.time_embedding(
+            sinusoidal_embedding_1d(self.freq_dim, t.flatten()).to(dtype=x[0].dtype))
+        e = e.reshape(t.shape[0], -1, e.shape[-1])
+        e0 = self.time_projection(e).unflatten(2, (6, self.dim))
+
+        # context
+        context = self.text_embedding(context)
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        for i, block in enumerate(self.blocks):
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"])
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(x, e=e0, freqs=freqs, context=context)
+            if audio_emb is not None:
+                x = self.audio_injector(x, i, audio_emb, audio_emb_global, seq_len)
+        # head
+        x = self.head(x, e)
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+        return x

comfy/ldm/wan/vae.py

+# original version: https://github.com/Wan-Video/Wan2.1/blob/main/wan/modules/vae.py
+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from comfy.ldm.modules.diffusionmodules.model import vae_attention
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+CACHE_T = 2
+
+
+class CausalConv3d(ops.Conv3d):
+    """
+    Causal 3d convolusion.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._padding = (self.padding[2], self.padding[2], self.padding[1],
+                         self.padding[1], 2 * self.padding[0], 0)
+        self.padding = (0, 0, 0)
+
+    def forward(self, x, cache_x=None, cache_list=None, cache_idx=None):
+        if cache_list is not None:
+            cache_x = cache_list[cache_idx]
+            cache_list[cache_idx] = None
+
+        padding = list(self._padding)
+        if cache_x is not None and self._padding[4] > 0:
+            cache_x = cache_x.to(x.device)
+            x = torch.cat([cache_x, x], dim=2)
+            padding[4] -= cache_x.shape[2]
+            del cache_x
+        x = F.pad(x, padding)
+
+        return super().forward(x)
+
+
+class RMS_norm(nn.Module):
+
+    def __init__(self, dim, channel_first=True, images=True, bias=False):
+        super().__init__()
+        broadcastable_dims = (1, 1, 1) if not images else (1, 1)
+        shape = (dim, *broadcastable_dims) if channel_first else (dim,)
+
+        self.channel_first = channel_first
+        self.scale = dim**0.5
+        self.gamma = nn.Parameter(torch.ones(shape))
+        self.bias = nn.Parameter(torch.zeros(shape)) if bias else None
+
+    # @torch.compile(mode="max-autotune-no-cudagraphs")
+    def forward(self, x):
+        return F.normalize(
+            x, dim=(1 if self.channel_first else -1)) * self.scale * self.gamma.to(x) + (self.bias.to(x) if self.bias is not None else 0)
+
+
+class Resample(nn.Module):
+
+    def __init__(self, dim, mode):
+        assert mode in ('none', 'upsample2d', 'upsample3d', 'downsample2d',
+                        'downsample3d')
+        super().__init__()
+        self.dim = dim
+        self.mode = mode
+
+        # layers
+        if mode == 'upsample2d':
+            self.resample = nn.Sequential(
+                nn.Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                ops.Conv2d(dim, dim // 2, 3, padding=1))
+        elif mode == 'upsample3d':
+            self.resample = nn.Sequential(
+                nn.Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                ops.Conv2d(dim, dim // 2, 3, padding=1))
+            self.time_conv = CausalConv3d(
+                dim, dim * 2, (3, 1, 1), padding=(1, 0, 0))
+
+        elif mode == 'downsample2d':
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                ops.Conv2d(dim, dim, 3, stride=(2, 2)))
+        elif mode == 'downsample3d':
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                ops.Conv2d(dim, dim, 3, stride=(2, 2)))
+            self.time_conv = CausalConv3d(
+                dim, dim, (3, 1, 1), stride=(2, 1, 1), padding=(0, 0, 0))
+
+        else:
+            self.resample = nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        b, c, t, h, w = x.size()
+        if self.mode == 'upsample3d':
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = 'Rep'
+                    feat_idx[0] += 1
+                else:
+
+                    cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                    if cache_x.shape[2] < 2 and feat_cache[
+                            idx] is not None and feat_cache[idx] != 'Rep':
+                        # cache last frame of last two chunk
+                        cache_x = torch.cat([
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device), cache_x
+                        ],
+                                            dim=2)
+                    if cache_x.shape[2] < 2 and feat_cache[
+                            idx] is not None and feat_cache[idx] == 'Rep':
+                        cache_x = torch.cat([
+                            torch.zeros_like(cache_x).to(cache_x.device),
+                            cache_x
+                        ],
+                                            dim=2)
+                    if feat_cache[idx] == 'Rep':
+                        x = self.time_conv(x)
+                    else:
+                        x = self.time_conv(x, feat_cache[idx])
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+
+                    x = x.reshape(b, 2, c, t, h, w)
+                    x = torch.stack((x[:, 0, :, :, :, :], x[:, 1, :, :, :, :]),
+                                    3)
+                    x = x.reshape(b, c, t * 2, h, w)
+        t = x.shape[2]
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+        x = self.resample(x)
+        x = rearrange(x, '(b t) c h w -> b c t h w', t=t)
+
+        if self.mode == 'downsample3d':
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = x.clone()
+                    feat_idx[0] += 1
+                else:
+
+                    cache_x = x[:, :, -1:, :, :].clone()
+                    # if cache_x.shape[2] < 2 and feat_cache[idx] is not None and feat_cache[idx]!='Rep':
+                    #     # cache last frame of last two chunk
+                    #     cache_x = torch.cat([feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device), cache_x], dim=2)
+
+                    x = self.time_conv(
+                        torch.cat([feat_cache[idx][:, :, -1:, :, :], x], 2))
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+        return x
+
+
+class ResidualBlock(nn.Module):
+
+    def __init__(self, in_dim, out_dim, dropout=0.0):
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+
+        # layers
+        self.residual = nn.Sequential(
+            RMS_norm(in_dim, images=False), nn.SiLU(),
+            CausalConv3d(in_dim, out_dim, 3, padding=1),
+            RMS_norm(out_dim, images=False), nn.SiLU(), nn.Dropout(dropout),
+            CausalConv3d(out_dim, out_dim, 3, padding=1))
+        self.shortcut = CausalConv3d(in_dim, out_dim, 1) \
+            if in_dim != out_dim else nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        old_x = x
+        for layer in self.residual:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, cache_list=feat_cache, cache_idx=idx)
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x + self.shortcut(old_x)
+
+
+class AttentionBlock(nn.Module):
+    """
+    Causal self-attention with a single head.
+    """
+
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+
+        # layers
+        self.norm = RMS_norm(dim)
+        self.to_qkv = ops.Conv2d(dim, dim * 3, 1)
+        self.proj = ops.Conv2d(dim, dim, 1)
+        self.optimized_attention = vae_attention()
+
+    def forward(self, x):
+        identity = x
+        b, c, t, h, w = x.size()
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+        x = self.norm(x)
+        # compute query, key, value
+
+        q, k, v = self.to_qkv(x).chunk(3, dim=1)
+        x = self.optimized_attention(q, k, v)
+
+        # output
+        x = self.proj(x)
+        x = rearrange(x, '(b t) c h w-> b c t h w', t=t)
+        return x + identity
+
+
+class Encoder3d(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[True, True, False],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+
+        # dimensions
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        # init block
+        self.conv1 = CausalConv3d(3, dims[0], 3, padding=1)
+
+        # downsample blocks
+        downsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            for _ in range(num_res_blocks):
+                downsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    downsamples.append(AttentionBlock(out_dim))
+                in_dim = out_dim
+
+            # downsample block
+            if i != len(dim_mult) - 1:
+                mode = 'downsample3d' if temperal_downsample[
+                    i] else 'downsample2d'
+                downsamples.append(Resample(out_dim, mode=mode))
+                scale /= 2.0
+        self.downsamples = nn.Sequential(*downsamples)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(out_dim, out_dim, dropout), AttentionBlock(out_dim),
+            ResidualBlock(out_dim, out_dim, dropout))
+
+        # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False), nn.SiLU(),
+            CausalConv3d(out_dim, z_dim, 3, padding=1))
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([
+                    feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                        cache_x.device), cache_x
+                ],
+                                    dim=2)
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## downsamples
+        for layer in self.downsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## middle
+        for layer in self.middle:
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+class Decoder3d(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_upsample=[False, True, True],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_upsample = temperal_upsample
+
+        # dimensions
+        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
+        scale = 1.0 / 2**(len(dim_mult) - 2)
+
+        # init block
+        self.conv1 = CausalConv3d(z_dim, dims[0], 3, padding=1)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(dims[0], dims[0], dropout), AttentionBlock(dims[0]),
+            ResidualBlock(dims[0], dims[0], dropout))
+
+        # upsample blocks
+        upsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            if i == 1 or i == 2 or i == 3:
+                in_dim = in_dim // 2
+            for _ in range(num_res_blocks + 1):
+                upsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    upsamples.append(AttentionBlock(out_dim))
+                in_dim = out_dim
+
+            # upsample block
+            if i != len(dim_mult) - 1:
+                mode = 'upsample3d' if temperal_upsample[i] else 'upsample2d'
+                upsamples.append(Resample(out_dim, mode=mode))
+                scale *= 2.0
+        self.upsamples = nn.Sequential(*upsamples)
+
+        # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False), nn.SiLU(),
+            CausalConv3d(out_dim, 3, 3, padding=1))
+        
+    # @torch.compile(mode="max-autotune-no-cudagraphs")
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        ## conv1
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([
+                    feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                        cache_x.device), cache_x
+                ],
+                                    dim=2)
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## middle
+        for layer in self.middle:
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## upsamples
+        for layer in self.upsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+def count_conv3d(model):
+    count = 0
+    for m in model.modules():
+        if isinstance(m, CausalConv3d):
+            count += 1
+    return count
+
+
+class WanVAE(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[True, True, False],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+        self.temperal_upsample = temperal_downsample[::-1]
+
+        # modules
+        self.encoder = Encoder3d(dim, z_dim * 2, dim_mult, num_res_blocks,
+                                 attn_scales, self.temperal_downsample, dropout)
+        self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
+        self.conv2 = CausalConv3d(z_dim, z_dim, 1)
+        self.decoder = Decoder3d(dim, z_dim, dim_mult, num_res_blocks,
+                                 attn_scales, self.temperal_upsample, dropout)
+
+    def encode(self, x):
+        self.clear_cache()
+        ## cache
+        t = x.shape[2]
+        iter_ = 1 + (t - 1) // 4
+        ## 对encode输入的x，按时间拆分为1、4、4、4....
+        for i in range(iter_):
+            self._enc_conv_idx = [0]
+            if i == 0:
+                out = self.encoder(
+                    x[:, :, :1, :, :],
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx)
+            else:
+                out_ = self.encoder(
+                    x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx)
+                out = torch.cat([out, out_], 2)
+        mu, log_var = self.conv1(out).chunk(2, dim=1)
+        self.clear_cache()
+        return mu
+
+    def decode(self, z):
+        self.clear_cache()
+        # z: [b,c,t,h,w]
+
+        iter_ = z.shape[2]
+        x = self.conv2(z)
+        for i in range(iter_):
+            self._conv_idx = [0]
+            if i == 0:
+                out = self.decoder(
+                    x[:, :, i:i + 1, :, :],
+                    feat_cache=self._feat_map,
+                    feat_idx=self._conv_idx)
+            else:
+                out_ = self.decoder(
+                    x[:, :, i:i + 1, :, :],
+                    feat_cache=self._feat_map,
+                    feat_idx=self._conv_idx)
+                out = torch.cat([out, out_], 2)
+        self.clear_cache()
+        return out
+
+    def clear_cache(self):
+        self._conv_num = count_conv3d(self.decoder)
+        self._conv_idx = [0]
+        self._feat_map = [None] * self._conv_num
+        #cache encode
+        self._enc_conv_num = count_conv3d(self.encoder)
+        self._enc_conv_idx = [0]
+        self._enc_feat_map = [None] * self._enc_conv_num

comfy/ldm/wan/vae2_2.py

+# original version: https://github.com/Wan-Video/Wan2.2/blob/main/wan/modules/vae2_2.py
+# Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from einops import rearrange
+from .vae import AttentionBlock, CausalConv3d, RMS_norm
+
+import comfy.ops
+ops = comfy.ops.disable_weight_init
+
+CACHE_T = 2
+
+
+class Resample(nn.Module):
+
+    def __init__(self, dim, mode):
+        assert mode in (
+            "none",
+            "upsample2d",
+            "upsample3d",
+            "downsample2d",
+            "downsample3d",
+        )
+        super().__init__()
+        self.dim = dim
+        self.mode = mode
+
+        # layers
+        if mode == "upsample2d":
+            self.resample = nn.Sequential(
+                nn.Upsample(scale_factor=(2.0, 2.0), mode="nearest-exact"),
+                ops.Conv2d(dim, dim, 3, padding=1),
+            )
+        elif mode == "upsample3d":
+            self.resample = nn.Sequential(
+                nn.Upsample(scale_factor=(2.0, 2.0), mode="nearest-exact"),
+                ops.Conv2d(dim, dim, 3, padding=1),
+                # ops.Conv2d(dim, dim//2, 3, padding=1)
+            )
+            self.time_conv = CausalConv3d(
+                dim, dim * 2, (3, 1, 1), padding=(1, 0, 0))
+        elif mode == "downsample2d":
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                ops.Conv2d(dim, dim, 3, stride=(2, 2)))
+        elif mode == "downsample3d":
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                ops.Conv2d(dim, dim, 3, stride=(2, 2)))
+            self.time_conv = CausalConv3d(
+                dim, dim, (3, 1, 1), stride=(2, 1, 1), padding=(0, 0, 0))
+        else:
+            self.resample = nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        b, c, t, h, w = x.size()
+        if self.mode == "upsample3d":
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = "Rep"
+                    feat_idx[0] += 1
+                else:
+                    cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                    if (cache_x.shape[2] < 2 and feat_cache[idx] is not None and
+                            feat_cache[idx] != "Rep"):
+                        # cache last frame of last two chunk
+                        cache_x = torch.cat(
+                            [
+                                feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                    cache_x.device),
+                                cache_x,
+                            ],
+                            dim=2,
+                        )
+                    if (cache_x.shape[2] < 2 and feat_cache[idx] is not None and
+                            feat_cache[idx] == "Rep"):
+                        cache_x = torch.cat(
+                            [
+                                torch.zeros_like(cache_x).to(cache_x.device),
+                                cache_x
+                            ],
+                            dim=2,
+                        )
+                    if feat_cache[idx] == "Rep":
+                        x = self.time_conv(x)
+                    else:
+                        x = self.time_conv(x, feat_cache[idx])
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+                    x = x.reshape(b, 2, c, t, h, w)
+                    x = torch.stack((x[:, 0, :, :, :, :], x[:, 1, :, :, :, :]),
+                                    3)
+                    x = x.reshape(b, c, t * 2, h, w)
+        t = x.shape[2]
+        x = rearrange(x, "b c t h w -> (b t) c h w")
+        x = self.resample(x)
+        x = rearrange(x, "(b t) c h w -> b c t h w", t=t)
+
+        if self.mode == "downsample3d":
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = x.clone()
+                    feat_idx[0] += 1
+                else:
+                    cache_x = x[:, :, -1:, :, :].clone()
+                    x = self.time_conv(
+                        torch.cat([feat_cache[idx][:, :, -1:, :, :], x], 2))
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+        return x
+
+
+class ResidualBlock(nn.Module):
+
+    def __init__(self, in_dim, out_dim, dropout=0.0):
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+
+        # layers
+        self.residual = nn.Sequential(
+            RMS_norm(in_dim, images=False),
+            nn.SiLU(),
+            CausalConv3d(in_dim, out_dim, 3, padding=1),
+            RMS_norm(out_dim, images=False),
+            nn.SiLU(),
+            nn.Dropout(dropout),
+            CausalConv3d(out_dim, out_dim, 3, padding=1),
+        )
+        self.shortcut = (
+            CausalConv3d(in_dim, out_dim, 1)
+            if in_dim != out_dim else nn.Identity())
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        old_x = x
+        for layer in self.residual:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat(
+                        [
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device),
+                            cache_x,
+                        ],
+                        dim=2,
+                    )
+                x = layer(x, cache_list=feat_cache, cache_idx=idx)
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x + self.shortcut(old_x)
+
+
+def patchify(x, patch_size):
+    if patch_size == 1:
+        return x
+    if x.dim() == 4:
+        x = rearrange(
+            x, "b c (h q) (w r) -> b (c r q) h w", q=patch_size, r=patch_size)
+    elif x.dim() == 5:
+        x = rearrange(
+            x,
+            "b c f (h q) (w r) -> b (c r q) f h w",
+            q=patch_size,
+            r=patch_size,
+        )
+    else:
+        raise ValueError(f"Invalid input shape: {x.shape}")
+
+    return x
+
+
+def unpatchify(x, patch_size):
+    if patch_size == 1:
+        return x
+
+    if x.dim() == 4:
+        x = rearrange(
+            x, "b (c r q) h w -> b c (h q) (w r)", q=patch_size, r=patch_size)
+    elif x.dim() == 5:
+        x = rearrange(
+            x,
+            "b (c r q) f h w -> b c f (h q) (w r)",
+            q=patch_size,
+            r=patch_size,
+        )
+    return x
+
+
+class AvgDown3D(nn.Module):
+
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        factor_t,
+        factor_s=1,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.factor_t = factor_t
+        self.factor_s = factor_s
+        self.factor = self.factor_t * self.factor_s * self.factor_s
+
+        assert in_channels * self.factor % out_channels == 0
+        self.group_size = in_channels * self.factor // out_channels
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        pad_t = (self.factor_t - x.shape[2] % self.factor_t) % self.factor_t
+        pad = (0, 0, 0, 0, pad_t, 0)
+        x = F.pad(x, pad)
+        B, C, T, H, W = x.shape
+        x = x.view(
+            B,
+            C,
+            T // self.factor_t,
+            self.factor_t,
+            H // self.factor_s,
+            self.factor_s,
+            W // self.factor_s,
+            self.factor_s,
+        )
+        x = x.permute(0, 1, 3, 5, 7, 2, 4, 6).contiguous()
+        x = x.view(
+            B,
+            C * self.factor,
+            T // self.factor_t,
+            H // self.factor_s,
+            W // self.factor_s,
+        )
+        x = x.view(
+            B,
+            self.out_channels,
+            self.group_size,
+            T // self.factor_t,
+            H // self.factor_s,
+            W // self.factor_s,
+        )
+        x = x.mean(dim=2)
+        return x
+
+
+class DupUp3D(nn.Module):
+
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        factor_t,
+        factor_s=1,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        self.factor_t = factor_t
+        self.factor_s = factor_s
+        self.factor = self.factor_t * self.factor_s * self.factor_s
+
+        assert out_channels * self.factor % in_channels == 0
+        self.repeats = out_channels * self.factor // in_channels
+
+    def forward(self, x: torch.Tensor, first_chunk=False) -> torch.Tensor:
+        x = x.repeat_interleave(self.repeats, dim=1)
+        x = x.view(
+            x.size(0),
+            self.out_channels,
+            self.factor_t,
+            self.factor_s,
+            self.factor_s,
+            x.size(2),
+            x.size(3),
+            x.size(4),
+        )
+        x = x.permute(0, 1, 5, 2, 6, 3, 7, 4).contiguous()
+        x = x.view(
+            x.size(0),
+            self.out_channels,
+            x.size(2) * self.factor_t,
+            x.size(4) * self.factor_s,
+            x.size(6) * self.factor_s,
+        )
+        if first_chunk:
+            x = x[:, :, self.factor_t - 1:, :, :]
+        return x
+
+
+class Down_ResidualBlock(nn.Module):
+
+    def __init__(self,
+                 in_dim,
+                 out_dim,
+                 dropout,
+                 mult,
+                 temperal_downsample=False,
+                 down_flag=False):
+        super().__init__()
+
+        # Shortcut path with downsample
+        self.avg_shortcut = AvgDown3D(
+            in_dim,
+            out_dim,
+            factor_t=2 if temperal_downsample else 1,
+            factor_s=2 if down_flag else 1,
+        )
+
+        # Main path with residual blocks and downsample
+        downsamples = []
+        for _ in range(mult):
+            downsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+            in_dim = out_dim
+
+        # Add the final downsample block
+        if down_flag:
+            mode = "downsample3d" if temperal_downsample else "downsample2d"
+            downsamples.append(Resample(out_dim, mode=mode))
+
+        self.downsamples = nn.Sequential(*downsamples)
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        x_copy = x
+        for module in self.downsamples:
+            x = module(x, feat_cache, feat_idx)
+
+        return x + self.avg_shortcut(x_copy)
+
+
+class Up_ResidualBlock(nn.Module):
+
+    def __init__(self,
+                 in_dim,
+                 out_dim,
+                 dropout,
+                 mult,
+                 temperal_upsample=False,
+                 up_flag=False):
+        super().__init__()
+        # Shortcut path with upsample
+        if up_flag:
+            self.avg_shortcut = DupUp3D(
+                in_dim,
+                out_dim,
+                factor_t=2 if temperal_upsample else 1,
+                factor_s=2 if up_flag else 1,
+            )
+        else:
+            self.avg_shortcut = None
+
+        # Main path with residual blocks and upsample
+        upsamples = []
+        for _ in range(mult):
+            upsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+            in_dim = out_dim
+
+        # Add the final upsample block
+        if up_flag:
+            mode = "upsample3d" if temperal_upsample else "upsample2d"
+            upsamples.append(Resample(out_dim, mode=mode))
+
+        self.upsamples = nn.Sequential(*upsamples)
+
+    def forward(self, x, feat_cache=None, feat_idx=[0], first_chunk=False):
+        x_main = x
+        for module in self.upsamples:
+            x_main = module(x_main, feat_cache, feat_idx)
+        if self.avg_shortcut is not None:
+            x_shortcut = self.avg_shortcut(x, first_chunk)
+            return x_main + x_shortcut
+        else:
+            return x_main
+
+
+class Encoder3d(nn.Module):
+
+    def __init__(
+        self,
+        dim=128,
+        z_dim=4,
+        dim_mult=[1, 2, 4, 4],
+        num_res_blocks=2,
+        attn_scales=[],
+        temperal_downsample=[True, True, False],
+        dropout=0.0,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+
+        # dimensions
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        # init block
+        self.conv1 = CausalConv3d(12, dims[0], 3, padding=1)
+
+        # downsample blocks
+        downsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            t_down_flag = (
+                temperal_downsample[i]
+                if i < len(temperal_downsample) else False)
+            downsamples.append(
+                Down_ResidualBlock(
+                    in_dim=in_dim,
+                    out_dim=out_dim,
+                    dropout=dropout,
+                    mult=num_res_blocks,
+                    temperal_downsample=t_down_flag,
+                    down_flag=i != len(dim_mult) - 1,
+                ))
+            scale /= 2.0
+        self.downsamples = nn.Sequential(*downsamples)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(out_dim, out_dim, dropout),
+            AttentionBlock(out_dim),
+            ResidualBlock(out_dim, out_dim, dropout),
+        )
+
+        # # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False),
+            nn.SiLU(),
+            CausalConv3d(out_dim, z_dim, 3, padding=1),
+        )
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device),
+                        cache_x,
+                    ],
+                    dim=2,
+                )
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## downsamples
+        for layer in self.downsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## middle
+        for layer in self.middle:
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    cache_x = torch.cat(
+                        [
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device),
+                            cache_x,
+                        ],
+                        dim=2,
+                    )
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+
+        return x
+
+
+class Decoder3d(nn.Module):
+
+    def __init__(
+        self,
+        dim=128,
+        z_dim=4,
+        dim_mult=[1, 2, 4, 4],
+        num_res_blocks=2,
+        attn_scales=[],
+        temperal_upsample=[False, True, True],
+        dropout=0.0,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_upsample = temperal_upsample
+
+        # dimensions
+        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
+        # init block
+        self.conv1 = CausalConv3d(z_dim, dims[0], 3, padding=1)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(dims[0], dims[0], dropout),
+            AttentionBlock(dims[0]),
+            ResidualBlock(dims[0], dims[0], dropout),
+        )
+
+        # upsample blocks
+        upsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            t_up_flag = temperal_upsample[i] if i < len(
+                temperal_upsample) else False
+            upsamples.append(
+                Up_ResidualBlock(
+                    in_dim=in_dim,
+                    out_dim=out_dim,
+                    dropout=dropout,
+                    mult=num_res_blocks + 1,
+                    temperal_upsample=t_up_flag,
+                    up_flag=i != len(dim_mult) - 1,
+                ))
+        self.upsamples = nn.Sequential(*upsamples)
+
+        # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False),
+            nn.SiLU(),
+            CausalConv3d(out_dim, 12, 3, padding=1),
+        )
+
+    def forward(self, x, feat_cache=None, feat_idx=[0], first_chunk=False):
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device),
+                        cache_x,
+                    ],
+                    dim=2,
+                )
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        for layer in self.middle:
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## upsamples
+        for layer in self.upsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx, first_chunk)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    cache_x = torch.cat(
+                        [
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device),
+                            cache_x,
+                        ],
+                        dim=2,
+                    )
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+def count_conv3d(model):
+    count = 0
+    for m in model.modules():
+        if isinstance(m, CausalConv3d):
+            count += 1
+    return count
+
+
+class WanVAE(nn.Module):
+
+    def __init__(
+        self,
+        dim=160,
+        dec_dim=256,
+        z_dim=16,
+        dim_mult=[1, 2, 4, 4],
+        num_res_blocks=2,
+        attn_scales=[],
+        temperal_downsample=[True, True, False],
+        dropout=0.0,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+        self.temperal_upsample = temperal_downsample[::-1]
+
+        # modules
+        self.encoder = Encoder3d(
+            dim,
+            z_dim * 2,
+            dim_mult,
+            num_res_blocks,
+            attn_scales,
+            self.temperal_downsample,
+            dropout,
+        )
+        self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
+        self.conv2 = CausalConv3d(z_dim, z_dim, 1)
+        self.decoder = Decoder3d(
+            dec_dim,
+            z_dim,
+            dim_mult,
+            num_res_blocks,
+            attn_scales,
+            self.temperal_upsample,
+            dropout,
+        )
+
+    def encode(self, x):
+        self.clear_cache()
+        x = patchify(x, patch_size=2)
+        t = x.shape[2]
+        iter_ = 1 + (t - 1) // 4
+        for i in range(iter_):
+            self._enc_conv_idx = [0]
+            if i == 0:
+                out = self.encoder(
+                    x[:, :, :1, :, :],
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx,
+                )
+            else:
+                out_ = self.encoder(
+                    x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
+                    feat_cache=self._enc_feat_map,
+                    feat_idx=self._enc_conv_idx,
+                )
+                out = torch.cat([out, out_], 2)
+        mu, log_var = self.conv1(out).chunk(2, dim=1)
+        self.clear_cache()
+        return mu
+
+    def decode(self, z):
+        self.clear_cache()
+        iter_ = z.shape[2]
+        x = self.conv2(z)
+        for i in range(iter_):
+            self._conv_idx = [0]
+            if i == 0:
+                out = self.decoder(
+                    x[:, :, i:i + 1, :, :],
+                    feat_cache=self._feat_map,
+                    feat_idx=self._conv_idx,
+                    first_chunk=True,
+                )
+            else:
+                out_ = self.decoder(
+                    x[:, :, i:i + 1, :, :],
+                    feat_cache=self._feat_map,
+                    feat_idx=self._conv_idx,
+                )
+                out = torch.cat([out, out_], 2)
+        out = unpatchify(out, patch_size=2)
+        self.clear_cache()
+        return out
+
+    def reparameterize(self, mu, log_var):
+        std = torch.exp(0.5 * log_var)
+        eps = torch.randn_like(std)
+        return eps * std + mu
+
+    def sample(self, imgs, deterministic=False):
+        mu, log_var = self.encode(imgs)
+        if deterministic:
+            return mu
+        std = torch.exp(0.5 * log_var.clamp(-30.0, 20.0))
+        return mu + std * torch.randn_like(std)
+
+    def clear_cache(self):
+        self._conv_num = count_conv3d(self.decoder)
+        self._conv_idx = [0]
+        self._feat_map = [None] * self._conv_num
+        # cache encode
+        self._enc_conv_num = count_conv3d(self.encoder)
+        self._enc_conv_idx = [0]
+        self._enc_feat_map = [None] * self._enc_conv_num

comfy/lora.py

+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Comfy
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
+from __future__ import annotations
+import comfy.utils
+import comfy.model_management
+import comfy.model_base
+import comfy.weight_adapter as weight_adapter
+import logging
+import torch
+
+LORA_CLIP_MAP = {
+    "mlp.fc1": "mlp_fc1",
+    "mlp.fc2": "mlp_fc2",
+    "self_attn.k_proj": "self_attn_k_proj",
+    "self_attn.q_proj": "self_attn_q_proj",
+    "self_attn.v_proj": "self_attn_v_proj",
+    "self_attn.out_proj": "self_attn_out_proj",
+}
+
+
+def load_lora(lora, to_load, log_missing=True):
+    patch_dict = {}
+    loaded_keys = set()
+    for x in to_load:
+        alpha_name = "{}.alpha".format(x)
+        alpha = None
+        if alpha_name in lora.keys():
+            alpha = lora[alpha_name].item()
+            loaded_keys.add(alpha_name)
+
+        dora_scale_name = "{}.dora_scale".format(x)
+        dora_scale = None
+        if dora_scale_name in lora.keys():
+            dora_scale = lora[dora_scale_name]
+            loaded_keys.add(dora_scale_name)
+
+        for adapter_cls in weight_adapter.adapters:
+            adapter = adapter_cls.load(x, lora, alpha, dora_scale, loaded_keys)
+            if adapter is not None:
+                patch_dict[to_load[x]] = adapter
+                loaded_keys.update(adapter.loaded_keys)
+                continue
+
+        w_norm_name = "{}.w_norm".format(x)
+        b_norm_name = "{}.b_norm".format(x)
+        w_norm = lora.get(w_norm_name, None)
+        b_norm = lora.get(b_norm_name, None)
+
+        if w_norm is not None:
+            loaded_keys.add(w_norm_name)
+            patch_dict[to_load[x]] = ("diff", (w_norm,))
+            if b_norm is not None:
+                loaded_keys.add(b_norm_name)
+                patch_dict["{}.bias".format(to_load[x][:-len(".weight")])] = ("diff", (b_norm,))
+
+        diff_name = "{}.diff".format(x)
+        diff_weight = lora.get(diff_name, None)
+        if diff_weight is not None:
+            patch_dict[to_load[x]] = ("diff", (diff_weight,))
+            loaded_keys.add(diff_name)
+
+        diff_bias_name = "{}.diff_b".format(x)
+        diff_bias = lora.get(diff_bias_name, None)
+        if diff_bias is not None:
+            patch_dict["{}.bias".format(to_load[x][:-len(".weight")])] = ("diff", (diff_bias,))
+            loaded_keys.add(diff_bias_name)
+
+        set_weight_name = "{}.set_weight".format(x)
+        set_weight = lora.get(set_weight_name, None)
+        if set_weight is not None:
+            patch_dict[to_load[x]] = ("set", (set_weight,))
+            loaded_keys.add(set_weight_name)
+
+    if log_missing:
+        for x in lora.keys():
+            if x not in loaded_keys:
+                logging.warning("lora key not loaded: {}".format(x))
+
+    return patch_dict
+
+def model_lora_keys_clip(model, key_map={}):
+    sdk = model.state_dict().keys()
+    for k in sdk:
+        if k.endswith(".weight"):
+            key_map["text_encoders.{}".format(k[:-len(".weight")])] = k #generic lora format without any weird key names
+
+    text_model_lora_key = "lora_te_text_model_encoder_layers_{}_{}"
+    clip_l_present = False
+    clip_g_present = False
+    for b in range(32): #TODO: clean up
+        for c in LORA_CLIP_MAP:
+            k = "clip_h.transformer.text_model.encoder.layers.{}.{}.weight".format(b, c)
+            if k in sdk:
+                lora_key = text_model_lora_key.format(b, LORA_CLIP_MAP[c])
+                key_map[lora_key] = k
+                lora_key = "lora_te1_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c])
+                key_map[lora_key] = k
+                lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
+                key_map[lora_key] = k
+
+            k = "clip_l.transformer.text_model.encoder.layers.{}.{}.weight".format(b, c)
+            if k in sdk:
+                lora_key = text_model_lora_key.format(b, LORA_CLIP_MAP[c])
+                key_map[lora_key] = k
+                lora_key = "lora_te1_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #SDXL base
+                key_map[lora_key] = k
+                clip_l_present = True
+                lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
+                key_map[lora_key] = k
+
+            k = "clip_g.transformer.text_model.encoder.layers.{}.{}.weight".format(b, c)
+            if k in sdk:
+                clip_g_present = True
+                if clip_l_present:
+                    lora_key = "lora_te2_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #SDXL base
+                    key_map[lora_key] = k
+                    lora_key = "text_encoder_2.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
+                    key_map[lora_key] = k
+                else:
+                    lora_key = "lora_te_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #TODO: test if this is correct for SDXL-Refiner
+                    key_map[lora_key] = k
+                    lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
+                    key_map[lora_key] = k
+                    lora_key = "lora_prior_te_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #cascade lora: TODO put lora key prefix in the model config
+                    key_map[lora_key] = k
+
+    for k in sdk:
+        if k.endswith(".weight"):
+            if k.startswith("t5xxl.transformer."):#OneTrainer SD3 and Flux lora
+                l_key = k[len("t5xxl.transformer."):-len(".weight")]
+                t5_index = 1
+                if clip_g_present:
+                    t5_index += 1
+                if clip_l_present:
+                    t5_index += 1
+                    if t5_index == 2:
+                        key_map["lora_te{}_{}".format(t5_index, l_key.replace(".", "_"))] = k #OneTrainer Flux
+                        t5_index += 1
+
+                key_map["lora_te{}_{}".format(t5_index, l_key.replace(".", "_"))] = k
+            elif k.startswith("hydit_clip.transformer.bert."): #HunyuanDiT Lora
+                l_key = k[len("hydit_clip.transformer.bert."):-len(".weight")]
+                lora_key = "lora_te1_{}".format(l_key.replace(".", "_"))
+                key_map[lora_key] = k
+
+
+    k = "clip_g.transformer.text_projection.weight"
+    if k in sdk:
+        key_map["lora_prior_te_text_projection"] = k #cascade lora?
+        # key_map["text_encoder.text_projection"] = k #TODO: check if other lora have the text_projection too
+        key_map["lora_te2_text_projection"] = k #OneTrainer SD3 lora
+
+    k = "clip_l.transformer.text_projection.weight"
+    if k in sdk:
+        key_map["lora_te1_text_projection"] = k #OneTrainer SD3 lora, not necessary but omits warning
+
+    return key_map
+
+def model_lora_keys_unet(model, key_map={}):
+    sd = model.state_dict()
+    sdk = sd.keys()
+
+    for k in sdk:
+        if k.startswith("diffusion_model."):
+            if k.endswith(".weight"):
+                key_lora = k[len("diffusion_model."):-len(".weight")].replace(".", "_")
+                key_map["lora_unet_{}".format(key_lora)] = k
+                key_map["{}".format(k[:-len(".weight")])] = k #generic lora format without any weird key names
+            else:
+                key_map["{}".format(k)] = k #generic lora format for not .weight without any weird key names
+
+    diffusers_keys = comfy.utils.unet_to_diffusers(model.model_config.unet_config)
+    for k in diffusers_keys:
+        if k.endswith(".weight"):
+            unet_key = "diffusion_model.{}".format(diffusers_keys[k])
+            key_lora = k[:-len(".weight")].replace(".", "_")
+            key_map["lora_unet_{}".format(key_lora)] = unet_key
+            key_map["lycoris_{}".format(key_lora)] = unet_key #simpletuner lycoris format
+
+            diffusers_lora_prefix = ["", "unet."]
+            for p in diffusers_lora_prefix:
+                diffusers_lora_key = "{}{}".format(p, k[:-len(".weight")].replace(".to_", ".processor.to_"))
+                if diffusers_lora_key.endswith(".to_out.0"):
+                    diffusers_lora_key = diffusers_lora_key[:-2]
+                key_map[diffusers_lora_key] = unet_key
+
+    if isinstance(model, comfy.model_base.StableCascade_C):
+        for k in sdk:
+            if k.startswith("diffusion_model."):
+                if k.endswith(".weight"):
+                    key_lora = k[len("diffusion_model."):-len(".weight")].replace(".", "_")
+                    key_map["lora_prior_unet_{}".format(key_lora)] = k
+
+    if isinstance(model, comfy.model_base.SD3): #Diffusers lora SD3
+        diffusers_keys = comfy.utils.mmdit_to_diffusers(model.model_config.unet_config, output_prefix="diffusion_model.")
+        for k in diffusers_keys:
+            if k.endswith(".weight"):
+                to = diffusers_keys[k]
+                key_lora = "transformer.{}".format(k[:-len(".weight")]) #regular diffusers sd3 lora format
+                key_map[key_lora] = to
+
+                key_lora = "base_model.model.{}".format(k[:-len(".weight")]) #format for flash-sd3 lora and others?
+                key_map[key_lora] = to
+
+                key_lora = "lora_transformer_{}".format(k[:-len(".weight")].replace(".", "_")) #OneTrainer lora
+                key_map[key_lora] = to
+
+                key_lora = "lycoris_{}".format(k[:-len(".weight")].replace(".", "_")) #simpletuner lycoris format
+                key_map[key_lora] = to
+
+    if isinstance(model, comfy.model_base.AuraFlow): #Diffusers lora AuraFlow
+        diffusers_keys = comfy.utils.auraflow_to_diffusers(model.model_config.unet_config, output_prefix="diffusion_model.")
+        for k in diffusers_keys:
+            if k.endswith(".weight"):
+                to = diffusers_keys[k]
+                key_lora = "transformer.{}".format(k[:-len(".weight")]) #simpletrainer and probably regular diffusers lora format
+                key_map[key_lora] = to
+
+    if isinstance(model, comfy.model_base.PixArt):
+        diffusers_keys = comfy.utils.pixart_to_diffusers(model.model_config.unet_config, output_prefix="diffusion_model.")
+        for k in diffusers_keys:
+            if k.endswith(".weight"):
+                to = diffusers_keys[k]
+                key_lora = "transformer.{}".format(k[:-len(".weight")]) #default format
+                key_map[key_lora] = to
+
+                key_lora = "base_model.model.{}".format(k[:-len(".weight")]) #diffusers training script
+                key_map[key_lora] = to
+
+                key_lora = "unet.base_model.model.{}".format(k[:-len(".weight")]) #old reference peft script
+                key_map[key_lora] = to
+
+    if isinstance(model, comfy.model_base.HunyuanDiT):
+        for k in sdk:
+            if k.startswith("diffusion_model.") and k.endswith(".weight"):
+                key_lora = k[len("diffusion_model."):-len(".weight")]
+                key_map["base_model.model.{}".format(key_lora)] = k #official hunyuan lora format
+
+    if isinstance(model, comfy.model_base.Flux): #Diffusers lora Flux
+        diffusers_keys = comfy.utils.flux_to_diffusers(model.model_config.unet_config, output_prefix="diffusion_model.")
+        for k in diffusers_keys:
+            if k.endswith(".weight"):
+                to = diffusers_keys[k]
+                key_map["transformer.{}".format(k[:-len(".weight")])] = to #simpletrainer and probably regular diffusers flux lora format
+                key_map["lycoris_{}".format(k[:-len(".weight")].replace(".", "_"))] = to #simpletrainer lycoris
+                key_map["lora_transformer_{}".format(k[:-len(".weight")].replace(".", "_"))] = to #onetrainer
+
+    if isinstance(model, comfy.model_base.GenmoMochi):
+        for k in sdk:
+            if k.startswith("diffusion_model.") and k.endswith(".weight"): #Official Mochi lora format
+                key_lora = k[len("diffusion_model."):-len(".weight")]
+                key_map["{}".format(key_lora)] = k
+
+    if isinstance(model, comfy.model_base.HunyuanVideo):
+        for k in sdk:
+            if k.startswith("diffusion_model.") and k.endswith(".weight"):
+                # diffusion-pipe lora format
+                key_lora = k
+                key_lora = key_lora.replace("_mod.lin.", "_mod.linear.").replace("_attn.qkv.", "_attn_qkv.").replace("_attn.proj.", "_attn_proj.")
+                key_lora = key_lora.replace("mlp.0.", "mlp.fc1.").replace("mlp.2.", "mlp.fc2.")
+                key_lora = key_lora.replace(".modulation.lin.", ".modulation.linear.")
+                key_lora = key_lora[len("diffusion_model."):-len(".weight")]
+                key_map["transformer.{}".format(key_lora)] = k
+                key_map["diffusion_model.{}".format(key_lora)] = k  # Old loras
+
+    if isinstance(model, comfy.model_base.HiDream):
+        for k in sdk:
+            if k.startswith("diffusion_model."):
+                if k.endswith(".weight"):
+                    key_lora = k[len("diffusion_model."):-len(".weight")]
+                    key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = k #SimpleTuner lycoris format
+                    key_map["transformer.{}".format(key_lora)] = k #SimpleTuner regular format
+
+    if isinstance(model, comfy.model_base.ACEStep):
+        for k in sdk:
+            if k.startswith("diffusion_model.") and k.endswith(".weight"): #Official ACE step lora format
+                key_lora = k[len("diffusion_model."):-len(".weight")]
+                key_map["{}".format(key_lora)] = k
+
+    if isinstance(model, comfy.model_base.QwenImage):
+        for k in sdk:
+            if k.startswith("diffusion_model.") and k.endswith(".weight"): #QwenImage lora format
+                key_lora = k[len("diffusion_model."):-len(".weight")]
+                # Direct mapping for transformer_blocks format (QwenImage LoRA format)
+                key_map["{}".format(key_lora)] = k
+                # Support transformer prefix format
+                key_map["transformer.{}".format(key_lora)] = k
+                key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = k #SimpleTuner lycoris format
+
+    return key_map
+
+
+def pad_tensor_to_shape(tensor: torch.Tensor, new_shape: list[int]) -> torch.Tensor:
+    """
+    Pad a tensor to a new shape with zeros.
+
+    Args:
+        tensor (torch.Tensor): The original tensor to be padded.
+        new_shape (List[int]): The desired shape of the padded tensor.
+
+    Returns:
+        torch.Tensor: A new tensor padded with zeros to the specified shape.
+
+    Note:
+        If the new shape is smaller than the original tensor in any dimension,
+        the original tensor will be truncated in that dimension.
+    """
+    if any([new_shape[i] < tensor.shape[i] for i in range(len(new_shape))]):
+        raise ValueError("The new shape must be larger than the original tensor in all dimensions")
+
+    if len(new_shape) != len(tensor.shape):
+        raise ValueError("The new shape must have the same number of dimensions as the original tensor")
+
+    # Create a new tensor filled with zeros
+    padded_tensor = torch.zeros(new_shape, dtype=tensor.dtype, device=tensor.device)
+
+    # Create slicing tuples for both tensors
+    orig_slices = tuple(slice(0, dim) for dim in tensor.shape)
+    new_slices = tuple(slice(0, dim) for dim in tensor.shape)
+
+    # Copy the original tensor into the new tensor
+    padded_tensor[new_slices] = tensor[orig_slices]
+
+    return padded_tensor
+
+def calculate_weight(patches, weight, key, intermediate_dtype=torch.float32, original_weights=None):
+    for p in patches:
+        strength = p[0]
+        v = p[1]
+        strength_model = p[2]
+        offset = p[3]
+        function = p[4]
+        if function is None:
+            function = lambda a: a
+
+        old_weight = None
+        if offset is not None:
+            old_weight = weight
+            weight = weight.narrow(offset[0], offset[1], offset[2])
+
+        if strength_model != 1.0:
+            weight *= strength_model
+
+        if isinstance(v, list):
+            v = (calculate_weight(v[1:], v[0][1](comfy.model_management.cast_to_device(v[0][0], weight.device, intermediate_dtype, copy=True), inplace=True), key, intermediate_dtype=intermediate_dtype), )
+
+        if isinstance(v, weight_adapter.WeightAdapterBase):
+            output = v.calculate_weight(weight, key, strength, strength_model, offset, function, intermediate_dtype, original_weights)
+            if output is None:
+                logging.warning("Calculate Weight Failed: {} {}".format(v.name, key))
+            else:
+                weight = output
+                if old_weight is not None:
+                    weight = old_weight
+            continue
+
+        if len(v) == 1:
+            patch_type = "diff"
+        elif len(v) == 2:
+            patch_type = v[0]
+            v = v[1]
+
+        if patch_type == "diff":
+            diff: torch.Tensor = v[0]
+            # An extra flag to pad the weight if the diff's shape is larger than the weight
+            do_pad_weight = len(v) > 1 and v[1]['pad_weight']
+            if do_pad_weight and diff.shape != weight.shape:
+                logging.info("Pad weight {} from {} to shape: {}".format(key, weight.shape, diff.shape))
+                weight = pad_tensor_to_shape(weight, diff.shape)
+
+            if strength != 0.0:
+                if diff.shape != weight.shape:
+                    logging.warning("WARNING SHAPE MISMATCH {} WEIGHT NOT MERGED {} != {}".format(key, diff.shape, weight.shape))
+                else:
+                    weight += function(strength * comfy.model_management.cast_to_device(diff, weight.device, weight.dtype))
+        elif patch_type == "set":
+            weight.copy_(v[0])
+        elif patch_type == "model_as_lora":
+            target_weight: torch.Tensor = v[0]
+            diff_weight = comfy.model_management.cast_to_device(target_weight, weight.device, intermediate_dtype) - \
+                          comfy.model_management.cast_to_device(original_weights[key][0][0], weight.device, intermediate_dtype)
+            weight += function(strength * comfy.model_management.cast_to_device(diff_weight, weight.device, weight.dtype))
+        else:
+            logging.warning("patch type not recognized {} {}".format(patch_type, key))
+
+        if old_weight is not None:
+            weight = old_weight
+
+    return weight

comfy/lora_convert.py

+import torch
+import comfy.utils
+
+
+def convert_lora_bfl_control(sd): #BFL loras for Flux
+    sd_out = {}
+    for k in sd:
+        k_to = "diffusion_model.{}".format(k.replace(".lora_B.bias", ".diff_b").replace("_norm.scale", "_norm.scale.set_weight"))
+        sd_out[k_to] = sd[k]
+
+    sd_out["diffusion_model.img_in.reshape_weight"] = torch.tensor([sd["img_in.lora_B.weight"].shape[0], sd["img_in.lora_A.weight"].shape[1]])
+    return sd_out
+
+
+def convert_lora_wan_fun(sd): #Wan Fun loras
+    return comfy.utils.state_dict_prefix_replace(sd, {"lora_unet__": "lora_unet_"})
+
+
+def convert_lora(sd):
+    if "img_in.lora_A.weight" in sd and "single_blocks.0.norm.key_norm.scale" in sd:
+        return convert_lora_bfl_control(sd)
+    if "lora_unet__blocks_0_cross_attn_k.lora_down.weight" in sd:
+        return convert_lora_wan_fun(sd)
+    return sd

comfy/model_base.py

+"""
+    This file is part of ComfyUI.
+    Copyright (C) 2024 Comfy
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+"""
+
+import torch
+import logging
+from comfy.ldm.modules.diffusionmodules.openaimodel import UNetModel, Timestep
+from comfy.ldm.cascade.stage_c import StageC
+from comfy.ldm.cascade.stage_b import StageB
+from comfy.ldm.modules.encoders.noise_aug_modules import CLIPEmbeddingNoiseAugmentation
+from comfy.ldm.modules.diffusionmodules.upscaling import ImageConcatWithNoiseAugmentation
+from comfy.ldm.modules.diffusionmodules.mmdit import OpenAISignatureMMDITWrapper
+import comfy.ldm.genmo.joint_model.asymm_models_joint
+import comfy.ldm.aura.mmdit
+import comfy.ldm.pixart.pixartms
+import comfy.ldm.hydit.models
+import comfy.ldm.audio.dit
+import comfy.ldm.audio.embedders
+import comfy.ldm.flux.model
+import comfy.ldm.lightricks.model
+import comfy.ldm.hunyuan_video.model
+import comfy.ldm.cosmos.model
+import comfy.ldm.cosmos.predict2
+import comfy.ldm.lumina.model
+import comfy.ldm.wan.model
+import comfy.ldm.hunyuan3d.model
+import comfy.ldm.hidream.model
+import comfy.ldm.chroma.model
+import comfy.ldm.ace.model
+import comfy.ldm.omnigen.omnigen2
+import comfy.ldm.qwen_image.model
+
+import comfy.model_management
+import comfy.patcher_extension
+import comfy.conds
+import comfy.ops
+from enum import Enum
+from . import utils
+import comfy.latent_formats
+import comfy.model_sampling
+import math
+from typing import TYPE_CHECKING
+if TYPE_CHECKING:
+    from comfy.model_patcher import ModelPatcher
+
+class ModelType(Enum):
+    EPS = 1
+    V_PREDICTION = 2
+    V_PREDICTION_EDM = 3
+    STABLE_CASCADE = 4
+    EDM = 5
+    FLOW = 6
+    V_PREDICTION_CONTINUOUS = 7
+    FLUX = 8
+    IMG_TO_IMG = 9
+    FLOW_COSMOS = 10
+
+
+def model_sampling(model_config, model_type):
+    s = comfy.model_sampling.ModelSamplingDiscrete
+
+    if model_type == ModelType.EPS:
+        c = comfy.model_sampling.EPS
+    elif model_type == ModelType.V_PREDICTION:
+        c = comfy.model_sampling.V_PREDICTION
+    elif model_type == ModelType.V_PREDICTION_EDM:
+        c = comfy.model_sampling.V_PREDICTION
+        s = comfy.model_sampling.ModelSamplingContinuousEDM
+    elif model_type == ModelType.FLOW:
+        c = comfy.model_sampling.CONST
+        s = comfy.model_sampling.ModelSamplingDiscreteFlow
+    elif model_type == ModelType.STABLE_CASCADE:
+        c = comfy.model_sampling.EPS
+        s = comfy.model_sampling.StableCascadeSampling
+    elif model_type == ModelType.EDM:
+        c = comfy.model_sampling.EDM
+        s = comfy.model_sampling.ModelSamplingContinuousEDM
+    elif model_type == ModelType.V_PREDICTION_CONTINUOUS:
+        c = comfy.model_sampling.V_PREDICTION
+        s = comfy.model_sampling.ModelSamplingContinuousV
+    elif model_type == ModelType.FLUX:
+        c = comfy.model_sampling.CONST
+        s = comfy.model_sampling.ModelSamplingFlux
+    elif model_type == ModelType.IMG_TO_IMG:
+        c = comfy.model_sampling.IMG_TO_IMG
+    elif model_type == ModelType.FLOW_COSMOS:
+        c = comfy.model_sampling.COSMOS_RFLOW
+        s = comfy.model_sampling.ModelSamplingCosmosRFlow
+
+    class ModelSampling(s, c):
+        pass
+
+    return ModelSampling(model_config)
+
+
+def convert_tensor(extra, dtype, device):
+    if hasattr(extra, "dtype"):
+        if extra.dtype != torch.int and extra.dtype != torch.long:
+            extra = comfy.model_management.cast_to_device(extra, device, dtype)
+        else:
+            extra = comfy.model_management.cast_to_device(extra, device, None)
+    return extra
+
+
+class BaseModel(torch.nn.Module):
+    def __init__(self, model_config, model_type=ModelType.EPS, device=None, unet_model=UNetModel):
+        super().__init__()
+
+        unet_config = model_config.unet_config
+        self.latent_format = model_config.latent_format
+        self.model_config = model_config
+        self.manual_cast_dtype = model_config.manual_cast_dtype
+        self.device = device
+        self.current_patcher: 'ModelPatcher' = None
+
+        if not unet_config.get("disable_unet_model_creation", False):
+            if model_config.custom_operations is None:
+                fp8 = model_config.optimizations.get("fp8", False)
+                operations = comfy.ops.pick_operations(unet_config.get("dtype", None), self.manual_cast_dtype, fp8_optimizations=fp8, scaled_fp8=model_config.scaled_fp8)
+            else:
+                operations = model_config.custom_operations
+            self.diffusion_model = unet_model(**unet_config, device=device, operations=operations)
+            if comfy.model_management.force_channels_last():
+                self.diffusion_model.to(memory_format=torch.channels_last)
+                logging.debug("using channels last mode for diffusion model")
+            logging.info("model weight dtype {}, manual cast: {}".format(self.get_dtype(), self.manual_cast_dtype))
+        self.model_type = model_type
+        self.model_sampling = model_sampling(model_config, model_type)
+
+        self.adm_channels = unet_config.get("adm_in_channels", None)
+        if self.adm_channels is None:
+            self.adm_channels = 0
+
+        self.concat_keys = ()
+        logging.info("model_type {}".format(model_type.name))
+        logging.debug("adm {}".format(self.adm_channels))
+        self.memory_usage_factor = model_config.memory_usage_factor
+        self.memory_usage_factor_conds = ()
+        self.memory_usage_shape_process = {}
+
+    def apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs):
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._apply_model,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.APPLY_MODEL, transformer_options)
+        ).execute(x, t, c_concat, c_crossattn, control, transformer_options, **kwargs)
+
+    def _apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, **kwargs):
+        sigma = t
+        xc = self.model_sampling.calculate_input(sigma, x)
+
+        if c_concat is not None:
+            xc = torch.cat([xc] + [comfy.model_management.cast_to_device(c_concat, xc.device, xc.dtype)], dim=1)
+
+        context = c_crossattn
+        dtype = self.get_dtype()
+
+        if self.manual_cast_dtype is not None:
+            dtype = self.manual_cast_dtype
+
+        xc = xc.to(dtype)
+        device = xc.device
+        t = self.model_sampling.timestep(t).float()
+        if context is not None:
+            context = comfy.model_management.cast_to_device(context, device, dtype)
+
+        extra_conds = {}
+        for o in kwargs:
+            extra = kwargs[o]
+
+            if hasattr(extra, "dtype"):
+                extra = convert_tensor(extra, dtype, device)
+            elif isinstance(extra, list):
+                ex = []
+                for ext in extra:
+                    ex.append(convert_tensor(ext, dtype, device))
+                extra = ex
+            extra_conds[o] = extra
+
+        t = self.process_timestep(t, x=x, **extra_conds)
+        model_output = self.diffusion_model(xc, t, context=context, control=control, transformer_options=transformer_options, **extra_conds).float()
+        return self.model_sampling.calculate_denoised(sigma, model_output, x)
+
+    def process_timestep(self, timestep, **kwargs):
+        return timestep
+
+    def get_dtype(self):
+        return self.diffusion_model.dtype
+
+    def encode_adm(self, **kwargs):
+        return None
+
+    def concat_cond(self, **kwargs):
+        if len(self.concat_keys) > 0:
+            cond_concat = []
+            denoise_mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+            concat_latent_image = kwargs.get("concat_latent_image", None)
+            if concat_latent_image is None:
+                concat_latent_image = kwargs.get("latent_image", None)
+            else:
+                concat_latent_image = self.process_latent_in(concat_latent_image)
+
+            noise = kwargs.get("noise", None)
+            device = kwargs["device"]
+
+            if concat_latent_image.shape[1:] != noise.shape[1:]:
+                concat_latent_image = utils.common_upscale(concat_latent_image, noise.shape[-1], noise.shape[-2], "bilinear", "center")
+                if noise.ndim == 5:
+                    if concat_latent_image.shape[-3] < noise.shape[-3]:
+                        concat_latent_image = torch.nn.functional.pad(concat_latent_image, (0, 0, 0, 0, 0, noise.shape[-3] - concat_latent_image.shape[-3]), "constant", 0)
+                    else:
+                        concat_latent_image = concat_latent_image[:, :, :noise.shape[-3]]
+
+            concat_latent_image = utils.resize_to_batch_size(concat_latent_image, noise.shape[0])
+
+            if denoise_mask is not None:
+                if len(denoise_mask.shape) == len(noise.shape):
+                    denoise_mask = denoise_mask[:, :1]
+
+                num_dim = noise.ndim - 2
+                denoise_mask = denoise_mask.reshape((-1, 1) + tuple(denoise_mask.shape[-num_dim:]))
+                if denoise_mask.shape[-2:] != noise.shape[-2:]:
+                    denoise_mask = utils.common_upscale(denoise_mask, noise.shape[-1], noise.shape[-2], "bilinear", "center")
+                denoise_mask = utils.resize_to_batch_size(denoise_mask.round(), noise.shape[0])
+
+            for ck in self.concat_keys:
+                if denoise_mask is not None:
+                    if ck == "mask":
+                        cond_concat.append(denoise_mask.to(device))
+                    elif ck == "masked_image":
+                        cond_concat.append(concat_latent_image.to(device))  # NOTE: the latent_image should be masked by the mask in pixel space
+                    elif ck == "mask_inverted":
+                        cond_concat.append(1.0 - denoise_mask.to(device))
+                else:
+                    if ck == "mask":
+                        cond_concat.append(torch.ones_like(noise)[:, :1])
+                    elif ck == "masked_image":
+                        cond_concat.append(self.blank_inpaint_image_like(noise))
+                    elif ck == "mask_inverted":
+                        cond_concat.append(torch.zeros_like(noise)[:, :1])
+                if ck == "concat_image":
+                    if concat_latent_image is not None:
+                        cond_concat.append(concat_latent_image.to(device))
+                    else:
+                        cond_concat.append(torch.zeros_like(noise))
+            data = torch.cat(cond_concat, dim=1)
+            return data
+        return None
+
+    def extra_conds(self, **kwargs):
+        out = {}
+        concat_cond = self.concat_cond(**kwargs)
+        if concat_cond is not None:
+            out['c_concat'] = comfy.conds.CONDNoiseShape(concat_cond)
+
+        adm = self.encode_adm(**kwargs)
+        if adm is not None:
+            out['y'] = comfy.conds.CONDRegular(adm)
+
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDCrossAttn(cross_attn)
+
+        cross_attn_cnet = kwargs.get("cross_attn_controlnet", None)
+        if cross_attn_cnet is not None:
+            out['crossattn_controlnet'] = comfy.conds.CONDCrossAttn(cross_attn_cnet)
+
+        c_concat = kwargs.get("noise_concat", None)
+        if c_concat is not None:
+            out['c_concat'] = comfy.conds.CONDNoiseShape(c_concat)
+
+        return out
+
+    def load_model_weights(self, sd, unet_prefix=""):
+        to_load = {}
+        keys = list(sd.keys())
+        for k in keys:
+            if k.startswith(unet_prefix):
+                to_load[k[len(unet_prefix):]] = sd.pop(k)
+
+        to_load = self.model_config.process_unet_state_dict(to_load)
+        m, u = self.diffusion_model.load_state_dict(to_load, strict=False)
+        if len(m) > 0:
+            logging.warning("unet missing: {}".format(m))
+
+        if len(u) > 0:
+            logging.warning("unet unexpected: {}".format(u))
+        del to_load
+        return self
+
+    def process_latent_in(self, latent):
+        return self.latent_format.process_in(latent)
+
+    def process_latent_out(self, latent):
+        return self.latent_format.process_out(latent)
+
+    def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None):
+        extra_sds = []
+        if clip_state_dict is not None:
+            extra_sds.append(self.model_config.process_clip_state_dict_for_saving(clip_state_dict))
+        if vae_state_dict is not None:
+            extra_sds.append(self.model_config.process_vae_state_dict_for_saving(vae_state_dict))
+        if clip_vision_state_dict is not None:
+            extra_sds.append(self.model_config.process_clip_vision_state_dict_for_saving(clip_vision_state_dict))
+
+        unet_state_dict = self.diffusion_model.state_dict()
+
+        if self.model_config.scaled_fp8 is not None:
+            unet_state_dict["scaled_fp8"] = torch.tensor([], dtype=self.model_config.scaled_fp8)
+
+        unet_state_dict = self.model_config.process_unet_state_dict_for_saving(unet_state_dict)
+
+        if self.model_type == ModelType.V_PREDICTION:
+            unet_state_dict["v_pred"] = torch.tensor([])
+
+        for sd in extra_sds:
+            unet_state_dict.update(sd)
+
+        return unet_state_dict
+
+    def set_inpaint(self):
+        self.concat_keys = ("mask", "masked_image")
+        def blank_inpaint_image_like(latent_image):
+            blank_image = torch.ones_like(latent_image)
+            # these are the values for "zero" in pixel space translated to latent space
+            blank_image[:,0] *= 0.8223
+            blank_image[:,1] *= -0.6876
+            blank_image[:,2] *= 0.6364
+            blank_image[:,3] *= 0.1380
+            return blank_image
+        self.blank_inpaint_image_like = blank_inpaint_image_like
+
+    def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
+        return self.model_sampling.noise_scaling(sigma.reshape([sigma.shape[0]] + [1] * (len(noise.shape) - 1)), noise, latent_image)
+
+    def memory_required(self, input_shape, cond_shapes={}):
+        input_shapes = [input_shape]
+        for c in self.memory_usage_factor_conds:
+            shape = cond_shapes.get(c, None)
+            if shape is not None:
+                if c in self.memory_usage_shape_process:
+                    out = []
+                    for s in shape:
+                        out.append(self.memory_usage_shape_process[c](s))
+                    shape = out
+
+                if len(shape) > 0:
+                    input_shapes += shape
+
+        if comfy.model_management.xformers_enabled() or comfy.model_management.pytorch_attention_flash_attention():
+            dtype = self.get_dtype()
+            if self.manual_cast_dtype is not None:
+                dtype = self.manual_cast_dtype
+            #TODO: this needs to be tweaked
+            area = sum(map(lambda input_shape: input_shape[0] * math.prod(input_shape[2:]), input_shapes))
+            return (area * comfy.model_management.dtype_size(dtype) * 0.01 * self.memory_usage_factor) * (1024 * 1024)
+        else:
+            #TODO: this formula might be too aggressive since I tweaked the sub-quad and split algorithms to use less memory.
+            area = sum(map(lambda input_shape: input_shape[0] * math.prod(input_shape[2:]), input_shapes))
+            return (area * 0.15 * self.memory_usage_factor) * (1024 * 1024)
+
+    def extra_conds_shapes(self, **kwargs):
+        return {}
+
+
+def unclip_adm(unclip_conditioning, device, noise_augmentor, noise_augment_merge=0.0, seed=None):
+    adm_inputs = []
+    weights = []
+    noise_aug = []
+    for unclip_cond in unclip_conditioning:
+        for adm_cond in unclip_cond["clip_vision_output"].image_embeds:
+            weight = unclip_cond["strength"]
+            noise_augment = unclip_cond["noise_augmentation"]
+            noise_level = round((noise_augmentor.max_noise_level - 1) * noise_augment)
+            c_adm, noise_level_emb = noise_augmentor(adm_cond.to(device), noise_level=torch.tensor([noise_level], device=device), seed=seed)
+            adm_out = torch.cat((c_adm, noise_level_emb), 1) * weight
+            weights.append(weight)
+            noise_aug.append(noise_augment)
+            adm_inputs.append(adm_out)
+
+    if len(noise_aug) > 1:
+        adm_out = torch.stack(adm_inputs).sum(0)
+        noise_augment = noise_augment_merge
+        noise_level = round((noise_augmentor.max_noise_level - 1) * noise_augment)
+        c_adm, noise_level_emb = noise_augmentor(adm_out[:, :noise_augmentor.time_embed.dim], noise_level=torch.tensor([noise_level], device=device))
+        adm_out = torch.cat((c_adm, noise_level_emb), 1)
+
+    return adm_out
+
+class SD21UNCLIP(BaseModel):
+    def __init__(self, model_config, noise_aug_config, model_type=ModelType.V_PREDICTION, device=None):
+        super().__init__(model_config, model_type, device=device)
+        self.noise_augmentor = CLIPEmbeddingNoiseAugmentation(**noise_aug_config)
+
+    def encode_adm(self, **kwargs):
+        unclip_conditioning = kwargs.get("unclip_conditioning", None)
+        device = kwargs["device"]
+        if unclip_conditioning is None:
+            return torch.zeros((1, self.adm_channels), device=device)
+        else:
+            return unclip_adm(unclip_conditioning, device, self.noise_augmentor, kwargs.get("unclip_noise_augment_merge", 0.05), kwargs.get("seed", 0) - 10)
+
+def sdxl_pooled(args, noise_augmentor):
+    if "unclip_conditioning" in args:
+        return unclip_adm(args.get("unclip_conditioning", None), args["device"], noise_augmentor, seed=args.get("seed", 0) - 10)[:,:1280]
+    else:
+        return args["pooled_output"]
+
+class SDXLRefiner(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.EPS, device=None):
+        super().__init__(model_config, model_type, device=device)
+        self.embedder = Timestep(256)
+        self.noise_augmentor = CLIPEmbeddingNoiseAugmentation(**{"noise_schedule_config": {"timesteps": 1000, "beta_schedule": "squaredcos_cap_v2"}, "timestep_dim": 1280})
+
+    def encode_adm(self, **kwargs):
+        clip_pooled = sdxl_pooled(kwargs, self.noise_augmentor)
+        width = kwargs.get("width", 768)
+        height = kwargs.get("height", 768)
+        crop_w = kwargs.get("crop_w", 0)
+        crop_h = kwargs.get("crop_h", 0)
+
+        if kwargs.get("prompt_type", "") == "negative":
+            aesthetic_score = kwargs.get("aesthetic_score", 2.5)
+        else:
+            aesthetic_score = kwargs.get("aesthetic_score", 6)
+
+        out = []
+        out.append(self.embedder(torch.Tensor([height])))
+        out.append(self.embedder(torch.Tensor([width])))
+        out.append(self.embedder(torch.Tensor([crop_h])))
+        out.append(self.embedder(torch.Tensor([crop_w])))
+        out.append(self.embedder(torch.Tensor([aesthetic_score])))
+        flat = torch.flatten(torch.cat(out)).unsqueeze(dim=0).repeat(clip_pooled.shape[0], 1)
+        return torch.cat((clip_pooled.to(flat.device), flat), dim=1)
+
+class SDXL(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.EPS, device=None):
+        super().__init__(model_config, model_type, device=device)
+        self.embedder = Timestep(256)
+        self.noise_augmentor = CLIPEmbeddingNoiseAugmentation(**{"noise_schedule_config": {"timesteps": 1000, "beta_schedule": "squaredcos_cap_v2"}, "timestep_dim": 1280})
+
+    def encode_adm(self, **kwargs):
+        clip_pooled = sdxl_pooled(kwargs, self.noise_augmentor)
+        width = kwargs.get("width", 768)
+        height = kwargs.get("height", 768)
+        crop_w = kwargs.get("crop_w", 0)
+        crop_h = kwargs.get("crop_h", 0)
+        target_width = kwargs.get("target_width", width)
+        target_height = kwargs.get("target_height", height)
+
+        out = []
+        out.append(self.embedder(torch.Tensor([height])))
+        out.append(self.embedder(torch.Tensor([width])))
+        out.append(self.embedder(torch.Tensor([crop_h])))
+        out.append(self.embedder(torch.Tensor([crop_w])))
+        out.append(self.embedder(torch.Tensor([target_height])))
+        out.append(self.embedder(torch.Tensor([target_width])))
+        flat = torch.flatten(torch.cat(out)).unsqueeze(dim=0).repeat(clip_pooled.shape[0], 1)
+        return torch.cat((clip_pooled.to(flat.device), flat), dim=1)
+
+
+class SVD_img2vid(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.V_PREDICTION_EDM, device=None):
+        super().__init__(model_config, model_type, device=device)
+        self.embedder = Timestep(256)
+
+    def encode_adm(self, **kwargs):
+        fps_id = kwargs.get("fps", 6) - 1
+        motion_bucket_id = kwargs.get("motion_bucket_id", 127)
+        augmentation = kwargs.get("augmentation_level", 0)
+
+        out = []
+        out.append(self.embedder(torch.Tensor([fps_id])))
+        out.append(self.embedder(torch.Tensor([motion_bucket_id])))
+        out.append(self.embedder(torch.Tensor([augmentation])))
+
+        flat = torch.flatten(torch.cat(out)).unsqueeze(dim=0)
+        return flat
+
+    def extra_conds(self, **kwargs):
+        out = {}
+        adm = self.encode_adm(**kwargs)
+        if adm is not None:
+            out['y'] = comfy.conds.CONDRegular(adm)
+
+        latent_image = kwargs.get("concat_latent_image", None)
+        noise = kwargs.get("noise", None)
+
+        if latent_image is None:
+            latent_image = torch.zeros_like(noise)
+
+        if latent_image.shape[1:] != noise.shape[1:]:
+            latent_image = utils.common_upscale(latent_image, noise.shape[-1], noise.shape[-2], "bilinear", "center")
+
+        latent_image = utils.resize_to_batch_size(latent_image, noise.shape[0])
+
+        out['c_concat'] = comfy.conds.CONDNoiseShape(latent_image)
+
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDCrossAttn(cross_attn)
+
+        if "time_conditioning" in kwargs:
+            out["time_context"] = comfy.conds.CONDCrossAttn(kwargs["time_conditioning"])
+
+        out['num_video_frames'] = comfy.conds.CONDConstant(noise.shape[0])
+        return out
+
+class SV3D_u(SVD_img2vid):
+    def encode_adm(self, **kwargs):
+        augmentation = kwargs.get("augmentation_level", 0)
+
+        out = []
+        out.append(self.embedder(torch.flatten(torch.Tensor([augmentation]))))
+
+        flat = torch.flatten(torch.cat(out)).unsqueeze(dim=0)
+        return flat
+
+class SV3D_p(SVD_img2vid):
+    def __init__(self, model_config, model_type=ModelType.V_PREDICTION_EDM, device=None):
+        super().__init__(model_config, model_type, device=device)
+        self.embedder_512 = Timestep(512)
+
+    def encode_adm(self, **kwargs):
+        augmentation = kwargs.get("augmentation_level", 0)
+        elevation = kwargs.get("elevation", 0) #elevation and azimuth are in degrees here
+        azimuth = kwargs.get("azimuth", 0)
+        noise = kwargs.get("noise", None)
+
+        out = []
+        out.append(self.embedder(torch.flatten(torch.Tensor([augmentation]))))
+        out.append(self.embedder_512(torch.deg2rad(torch.fmod(torch.flatten(90 - torch.Tensor([elevation])), 360.0))))
+        out.append(self.embedder_512(torch.deg2rad(torch.fmod(torch.flatten(torch.Tensor([azimuth])), 360.0))))
+
+        out = list(map(lambda a: utils.resize_to_batch_size(a, noise.shape[0]), out))
+        return torch.cat(out, dim=1)
+
+
+class Stable_Zero123(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.EPS, device=None, cc_projection_weight=None, cc_projection_bias=None):
+        super().__init__(model_config, model_type, device=device)
+        self.cc_projection = comfy.ops.manual_cast.Linear(cc_projection_weight.shape[1], cc_projection_weight.shape[0], dtype=self.get_dtype(), device=device)
+        self.cc_projection.weight.copy_(cc_projection_weight)
+        self.cc_projection.bias.copy_(cc_projection_bias)
+
+    def extra_conds(self, **kwargs):
+        out = {}
+
+        latent_image = kwargs.get("concat_latent_image", None)
+        noise = kwargs.get("noise", None)
+
+        if latent_image is None:
+            latent_image = torch.zeros_like(noise)
+
+        if latent_image.shape[1:] != noise.shape[1:]:
+            latent_image = utils.common_upscale(latent_image, noise.shape[-1], noise.shape[-2], "bilinear", "center")
+
+        latent_image = utils.resize_to_batch_size(latent_image, noise.shape[0])
+
+        out['c_concat'] = comfy.conds.CONDNoiseShape(latent_image)
+
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            if cross_attn.shape[-1] != 768:
+                cross_attn = self.cc_projection(cross_attn)
+            out['c_crossattn'] = comfy.conds.CONDCrossAttn(cross_attn)
+        return out
+
+class SD_X4Upscaler(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.V_PREDICTION, device=None):
+        super().__init__(model_config, model_type, device=device)
+        self.noise_augmentor = ImageConcatWithNoiseAugmentation(noise_schedule_config={"linear_start": 0.0001, "linear_end": 0.02}, max_noise_level=350)
+
+    def extra_conds(self, **kwargs):
+        out = {}
+
+        image = kwargs.get("concat_image", None)
+        noise = kwargs.get("noise", None)
+        noise_augment = kwargs.get("noise_augmentation", 0.0)
+        device = kwargs["device"]
+        seed = kwargs["seed"] - 10
+
+        noise_level = round((self.noise_augmentor.max_noise_level) * noise_augment)
+
+        if image is None:
+            image = torch.zeros_like(noise)[:,:3]
+
+        if image.shape[1:] != noise.shape[1:]:
+            image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+
+        noise_level = torch.tensor([noise_level], device=device)
+        if noise_augment > 0:
+            image, noise_level = self.noise_augmentor(image.to(device), noise_level=noise_level, seed=seed)
+
+        image = utils.resize_to_batch_size(image, noise.shape[0])
+
+        out['c_concat'] = comfy.conds.CONDNoiseShape(image)
+        out['y'] = comfy.conds.CONDRegular(noise_level)
+
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDCrossAttn(cross_attn)
+        return out
+
+class IP2P:
+    def concat_cond(self, **kwargs):
+        image = kwargs.get("concat_latent_image", None)
+        noise = kwargs.get("noise", None)
+        device = kwargs["device"]
+
+        if image is None:
+            image = torch.zeros_like(noise)
+        else:
+            image = image.to(device=device)
+
+        if image.shape[1:] != noise.shape[1:]:
+            image = utils.common_upscale(image, noise.shape[-1], noise.shape[-2], "bilinear", "center")
+
+        image = utils.resize_to_batch_size(image, noise.shape[0])
+        return self.process_ip2p_image_in(image)
+
+
+class SD15_instructpix2pix(IP2P, BaseModel):
+    def __init__(self, model_config, model_type=ModelType.EPS, device=None):
+        super().__init__(model_config, model_type, device=device)
+        self.process_ip2p_image_in = lambda image: image
+
+
+class SDXL_instructpix2pix(IP2P, SDXL):
+    def __init__(self, model_config, model_type=ModelType.EPS, device=None):
+        super().__init__(model_config, model_type, device=device)
+        if model_type == ModelType.V_PREDICTION_EDM:
+            self.process_ip2p_image_in = lambda image: comfy.latent_formats.SDXL().process_in(image) #cosxl ip2p
+        else:
+            self.process_ip2p_image_in = lambda image: image #diffusers ip2p
+
+class Lotus(BaseModel):
+    def extra_conds(self, **kwargs):
+        out = {}
+        cross_attn = kwargs.get("cross_attn", None)
+        out['c_crossattn'] = comfy.conds.CONDCrossAttn(cross_attn)
+        device = kwargs["device"]
+        task_emb = torch.tensor([1, 0]).float().to(device)
+        task_emb = torch.cat([torch.sin(task_emb), torch.cos(task_emb)]).unsqueeze(0)
+        out['y'] = comfy.conds.CONDRegular(task_emb)
+        return out
+
+    def __init__(self, model_config, model_type=ModelType.IMG_TO_IMG, device=None):
+        super().__init__(model_config, model_type, device=device)
+
+class StableCascade_C(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.STABLE_CASCADE, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=StageC)
+        self.diffusion_model.eval().requires_grad_(False)
+
+    def extra_conds(self, **kwargs):
+        out = {}
+        clip_text_pooled = kwargs["pooled_output"]
+        if clip_text_pooled is not None:
+            out['clip_text_pooled'] = comfy.conds.CONDRegular(clip_text_pooled)
+
+        if "unclip_conditioning" in kwargs:
+            embeds = []
+            for unclip_cond in kwargs["unclip_conditioning"]:
+                weight = unclip_cond["strength"]
+                embeds.append(unclip_cond["clip_vision_output"].image_embeds.unsqueeze(0) * weight)
+            clip_img = torch.cat(embeds, dim=1)
+        else:
+            clip_img = torch.zeros((1, 1, 768))
+        out["clip_img"] = comfy.conds.CONDRegular(clip_img)
+        out["sca"] = comfy.conds.CONDRegular(torch.zeros((1,)))
+        out["crp"] = comfy.conds.CONDRegular(torch.zeros((1,)))
+
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['clip_text'] = comfy.conds.CONDCrossAttn(cross_attn)
+        return out
+
+
+class StableCascade_B(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.STABLE_CASCADE, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=StageB)
+        self.diffusion_model.eval().requires_grad_(False)
+
+    def extra_conds(self, **kwargs):
+        out = {}
+        noise = kwargs.get("noise", None)
+
+        clip_text_pooled = kwargs["pooled_output"]
+        if clip_text_pooled is not None:
+            out['clip'] = comfy.conds.CONDRegular(clip_text_pooled)
+
+        #size of prior doesn't really matter if zeros because it gets resized but I still want it to get batched
+        prior = kwargs.get("stable_cascade_prior", torch.zeros((1, 16, (noise.shape[2] * 4) // 42, (noise.shape[3] * 4) // 42), dtype=noise.dtype, layout=noise.layout, device=noise.device))
+
+        out["effnet"] = comfy.conds.CONDRegular(prior.to(device=noise.device))
+        out["sca"] = comfy.conds.CONDRegular(torch.zeros((1,)))
+        return out
+
+
+class SD3(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=OpenAISignatureMMDITWrapper)
+
+    def encode_adm(self, **kwargs):
+        return kwargs["pooled_output"]
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        return out
+
+
+class AuraFlow(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.aura.mmdit.MMDiT)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        return out
+
+
+class StableAudio1(BaseModel):
+    def __init__(self, model_config, seconds_start_embedder_weights, seconds_total_embedder_weights, model_type=ModelType.V_PREDICTION_CONTINUOUS, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.audio.dit.AudioDiffusionTransformer)
+        self.seconds_start_embedder = comfy.ldm.audio.embedders.NumberConditioner(768, min_val=0, max_val=512)
+        self.seconds_total_embedder = comfy.ldm.audio.embedders.NumberConditioner(768, min_val=0, max_val=512)
+        self.seconds_start_embedder.load_state_dict(seconds_start_embedder_weights)
+        self.seconds_total_embedder.load_state_dict(seconds_total_embedder_weights)
+
+    def extra_conds(self, **kwargs):
+        out = {}
+
+        noise = kwargs.get("noise", None)
+        device = kwargs["device"]
+
+        seconds_start = kwargs.get("seconds_start", 0)
+        seconds_total = kwargs.get("seconds_total", int(noise.shape[-1] / 21.53))
+
+        seconds_start_embed = self.seconds_start_embedder([seconds_start])[0].to(device)
+        seconds_total_embed = self.seconds_total_embedder([seconds_total])[0].to(device)
+
+        global_embed = torch.cat([seconds_start_embed, seconds_total_embed], dim=-1).reshape((1, -1))
+        out['global_embed'] = comfy.conds.CONDRegular(global_embed)
+
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            cross_attn = torch.cat([cross_attn.to(device), seconds_start_embed.repeat((cross_attn.shape[0], 1, 1)), seconds_total_embed.repeat((cross_attn.shape[0], 1, 1))], dim=1)
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        return out
+
+    def state_dict_for_saving(self, clip_state_dict=None, vae_state_dict=None, clip_vision_state_dict=None):
+        sd = super().state_dict_for_saving(clip_state_dict=clip_state_dict, vae_state_dict=vae_state_dict, clip_vision_state_dict=clip_vision_state_dict)
+        d = {"conditioner.conditioners.seconds_start.": self.seconds_start_embedder.state_dict(), "conditioner.conditioners.seconds_total.": self.seconds_total_embedder.state_dict()}
+        for k in d:
+            s = d[k]
+            for l in s:
+                sd["{}{}".format(k, l)] = s[l]
+        return sd
+
+
+class HunyuanDiT(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.V_PREDICTION, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hydit.models.HunYuanDiT)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            out['text_embedding_mask'] = comfy.conds.CONDRegular(attention_mask)
+
+        conditioning_mt5xl = kwargs.get("conditioning_mt5xl", None)
+        if conditioning_mt5xl is not None:
+            out['encoder_hidden_states_t5'] = comfy.conds.CONDRegular(conditioning_mt5xl)
+
+        attention_mask_mt5xl = kwargs.get("attention_mask_mt5xl", None)
+        if attention_mask_mt5xl is not None:
+            out['text_embedding_mask_t5'] = comfy.conds.CONDRegular(attention_mask_mt5xl)
+
+        width = kwargs.get("width", 768)
+        height = kwargs.get("height", 768)
+        target_width = kwargs.get("target_width", width)
+        target_height = kwargs.get("target_height", height)
+
+        out['image_meta_size'] = comfy.conds.CONDRegular(torch.FloatTensor([[height, width, target_height, target_width, 0, 0]]))
+        return out
+
+class PixArt(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.EPS, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.pixart.pixartms.PixArtMS)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        width = kwargs.get("width", None)
+        height = kwargs.get("height", None)
+        if width is not None and height is not None:
+            out["c_size"] = comfy.conds.CONDRegular(torch.FloatTensor([[height, width]]))
+            out["c_ar"] = comfy.conds.CONDRegular(torch.FloatTensor([[kwargs.get("aspect_ratio", height/width)]]))
+
+        return out
+
+class Flux(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLUX, device=None, unet_model=comfy.ldm.flux.model.Flux):
+        super().__init__(model_config, model_type, device=device, unet_model=unet_model)
+        self.memory_usage_factor_conds = ("ref_latents",)
+
+    def concat_cond(self, **kwargs):
+        try:
+            #Handle Flux control loras dynamically changing the img_in weight.
+            num_channels = self.diffusion_model.img_in.weight.shape[1] // (self.diffusion_model.patch_size * self.diffusion_model.patch_size)
+        except:
+            #Some cases like tensorrt might not have the weights accessible
+            num_channels = self.model_config.unet_config["in_channels"]
+
+        out_channels = self.model_config.unet_config["out_channels"]
+
+        if num_channels <= out_channels:
+            return None
+
+        image = kwargs.get("concat_latent_image", None)
+        noise = kwargs.get("noise", None)
+        device = kwargs["device"]
+
+        if image is None:
+            image = torch.zeros_like(noise)
+
+        image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+        image = utils.resize_to_batch_size(image, noise.shape[0])
+        image = self.process_latent_in(image)
+        if num_channels <= out_channels * 2:
+            return image
+
+        #inpaint model
+        mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+        if mask is None:
+            mask = torch.ones_like(noise)[:, :1]
+
+        mask = torch.mean(mask, dim=1, keepdim=True)
+        mask = utils.common_upscale(mask.to(device), noise.shape[-1] * 8, noise.shape[-2] * 8, "bilinear", "center")
+        mask = mask.view(mask.shape[0], mask.shape[2] // 8, 8, mask.shape[3] // 8, 8).permute(0, 2, 4, 1, 3).reshape(mask.shape[0], -1, mask.shape[2] // 8, mask.shape[3] // 8)
+        mask = utils.resize_to_batch_size(mask, noise.shape[0])
+        return torch.cat((image, mask), dim=1)
+
+    def encode_adm(self, **kwargs):
+        return kwargs["pooled_output"]
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        # upscale the attention mask, since now we
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            shape = kwargs["noise"].shape
+            mask_ref_size = kwargs["attention_mask_img_shape"]
+            # the model will pad to the patch size, and then divide
+            # essentially dividing and rounding up
+            (h_tok, w_tok) = (math.ceil(shape[2] / self.diffusion_model.patch_size), math.ceil(shape[3] / self.diffusion_model.patch_size))
+            attention_mask = utils.upscale_dit_mask(attention_mask, mask_ref_size, (h_tok, w_tok))
+            out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
+
+        guidance = kwargs.get("guidance", 3.5)
+        if guidance is not None:
+            out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
+
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            latents = []
+            for lat in ref_latents:
+                latents.append(self.process_latent_in(lat))
+            out['ref_latents'] = comfy.conds.CONDList(latents)
+
+            ref_latents_method = kwargs.get("reference_latents_method", None)
+            if ref_latents_method is not None:
+                out['ref_latents_method'] = comfy.conds.CONDConstant(ref_latents_method)
+        return out
+
+    def extra_conds_shapes(self, **kwargs):
+        out = {}
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16])
+        return out
+
+
+class GenmoMochi(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.genmo.joint_model.asymm_models_joint.AsymmDiTJoint)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
+            out['num_tokens'] = comfy.conds.CONDConstant(max(1, torch.sum(attention_mask).item()))
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        return out
+
+class LTXV(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lightricks.model.LTXVModel) #TODO
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        out['frame_rate'] = comfy.conds.CONDConstant(kwargs.get("frame_rate", 25))
+
+        denoise_mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+        if denoise_mask is not None:
+            out["denoise_mask"] = comfy.conds.CONDRegular(denoise_mask)
+
+        keyframe_idxs = kwargs.get("keyframe_idxs", None)
+        if keyframe_idxs is not None:
+            out['keyframe_idxs'] = comfy.conds.CONDRegular(keyframe_idxs)
+
+        return out
+
+    def process_timestep(self, timestep, x, denoise_mask=None, **kwargs):
+        if denoise_mask is None:
+            return timestep
+        return self.diffusion_model.patchifier.patchify(((denoise_mask) * timestep.view([timestep.shape[0]] + [1] * (denoise_mask.ndim - 1)))[:, :1])[0]
+
+    def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
+        return latent_image
+
+class HunyuanVideo(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan_video.model.HunyuanVideo)
+
+    def encode_adm(self, **kwargs):
+        return kwargs["pooled_output"]
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        guidance = kwargs.get("guidance", 6.0)
+        if guidance is not None:
+            out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
+
+        guiding_frame_index = kwargs.get("guiding_frame_index", None)
+        if guiding_frame_index is not None:
+            out['guiding_frame_index'] = comfy.conds.CONDRegular(torch.FloatTensor([guiding_frame_index]))
+
+        ref_latent = kwargs.get("ref_latent", None)
+        if ref_latent is not None:
+            out['ref_latent'] = comfy.conds.CONDRegular(self.process_latent_in(ref_latent))
+
+        return out
+
+    def scale_latent_inpaint(self, latent_image, **kwargs):
+        return latent_image
+
+class HunyuanVideoI2V(HunyuanVideo):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device)
+        self.concat_keys = ("concat_image", "mask_inverted")
+
+    def scale_latent_inpaint(self, latent_image, **kwargs):
+        return super().scale_latent_inpaint(latent_image=latent_image, **kwargs)
+
+class HunyuanVideoSkyreelsI2V(HunyuanVideo):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device)
+        self.concat_keys = ("concat_image",)
+
+    def scale_latent_inpaint(self, latent_image, **kwargs):
+        return super().scale_latent_inpaint(latent_image=latent_image, **kwargs)
+
+class CosmosVideo(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.EDM, image_to_video=False, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.cosmos.model.GeneralDIT)
+        self.image_to_video = image_to_video
+        if self.image_to_video:
+            self.concat_keys = ("mask_inverted",)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        out['fps'] = comfy.conds.CONDConstant(kwargs.get("frame_rate", None))
+        return out
+
+    def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
+        sigma = sigma.reshape([sigma.shape[0]] + [1] * (len(noise.shape) - 1))
+        sigma_noise_augmentation = 0 #TODO
+        if sigma_noise_augmentation != 0:
+            latent_image = latent_image + noise
+        latent_image = self.model_sampling.calculate_input(torch.tensor([sigma_noise_augmentation], device=latent_image.device, dtype=latent_image.dtype), latent_image)
+        return latent_image * ((sigma ** 2 + self.model_sampling.sigma_data ** 2) ** 0.5)
+
+class CosmosPredict2(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW_COSMOS, image_to_video=False, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.cosmos.predict2.MiniTrainDIT)
+        self.image_to_video = image_to_video
+        if self.image_to_video:
+            self.concat_keys = ("mask_inverted",)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        denoise_mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+        if denoise_mask is not None:
+            out["denoise_mask"] = comfy.conds.CONDRegular(denoise_mask)
+
+        out['fps'] = comfy.conds.CONDConstant(kwargs.get("frame_rate", None))
+        return out
+
+    def process_timestep(self, timestep, x, denoise_mask=None, **kwargs):
+        if denoise_mask is None:
+            return timestep
+        if denoise_mask.ndim <= 4:
+            return timestep
+        condition_video_mask_B_1_T_1_1 = denoise_mask.mean(dim=[1, 3, 4], keepdim=True)
+        c_noise_B_1_T_1_1 = 0.0 * (1.0 - condition_video_mask_B_1_T_1_1) + timestep.reshape(timestep.shape[0], 1, 1, 1, 1) * condition_video_mask_B_1_T_1_1
+        out = c_noise_B_1_T_1_1.squeeze(dim=[1, 3, 4])
+        return out
+
+    def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
+        sigma = sigma.reshape([sigma.shape[0]] + [1] * (len(noise.shape) - 1))
+        sigma_noise_augmentation = 0 #TODO
+        if sigma_noise_augmentation != 0:
+            latent_image = latent_image + noise
+        latent_image = self.model_sampling.calculate_input(torch.tensor([sigma_noise_augmentation], device=latent_image.device, dtype=latent_image.dtype), latent_image)
+        sigma = (sigma / (sigma + 1))
+        return latent_image / (1.0 - sigma)
+
+class Lumina2(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.lumina.model.NextDiT)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            if torch.numel(attention_mask) != attention_mask.sum():
+                out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
+            out['num_tokens'] = comfy.conds.CONDConstant(max(1, torch.sum(attention_mask).item()))
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        return out
+
+class WAN21(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)
+        self.image_to_video = image_to_video
+
+    def concat_cond(self, **kwargs):
+        noise = kwargs.get("noise", None)
+        extra_channels = self.diffusion_model.patch_embedding.weight.shape[1] - noise.shape[1]
+        if extra_channels == 0:
+            return None
+
+        image = kwargs.get("concat_latent_image", None)
+        device = kwargs["device"]
+
+        if image is None:
+            shape_image = list(noise.shape)
+            shape_image[1] = extra_channels
+            image = torch.zeros(shape_image, dtype=noise.dtype, layout=noise.layout, device=noise.device)
+        else:
+            latent_dim = self.latent_format.latent_channels
+            image = utils.common_upscale(image.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+            for i in range(0, image.shape[1], latent_dim):
+                image[:, i: i + latent_dim] = self.process_latent_in(image[:, i: i + latent_dim])
+            image = utils.resize_to_batch_size(image, noise.shape[0])
+
+        if extra_channels != image.shape[1] + 4:
+            if not self.image_to_video or extra_channels == image.shape[1]:
+                return image
+
+        if image.shape[1] > (extra_channels - 4):
+            image = image[:, :(extra_channels - 4)]
+
+        mask = kwargs.get("concat_mask", kwargs.get("denoise_mask", None))
+        if mask is None:
+            mask = torch.zeros_like(noise)[:, :4]
+        else:
+            if mask.shape[1] != 4:
+                mask = torch.mean(mask, dim=1, keepdim=True)
+            mask = 1.0 - mask
+            mask = utils.common_upscale(mask.to(device), noise.shape[-1], noise.shape[-2], "bilinear", "center")
+            if mask.shape[-3] < noise.shape[-3]:
+                mask = torch.nn.functional.pad(mask, (0, 0, 0, 0, 0, noise.shape[-3] - mask.shape[-3]), mode='constant', value=0)
+            if mask.shape[1] == 1:
+                mask = mask.repeat(1, 4, 1, 1, 1)
+            mask = utils.resize_to_batch_size(mask, noise.shape[0])
+
+        concat_mask_index = kwargs.get("concat_mask_index", 0)
+        if concat_mask_index != 0:
+            return torch.cat((image[:, :concat_mask_index], mask, image[:, concat_mask_index:]), dim=1)
+        else:
+            return torch.cat((mask, image), dim=1)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        clip_vision_output = kwargs.get("clip_vision_output", None)
+        if clip_vision_output is not None:
+            out['clip_fea'] = comfy.conds.CONDRegular(clip_vision_output.penultimate_hidden_states)
+
+        time_dim_concat = kwargs.get("time_dim_concat", None)
+        if time_dim_concat is not None:
+            out['time_dim_concat'] = comfy.conds.CONDRegular(self.process_latent_in(time_dim_concat))
+
+        reference_latents = kwargs.get("reference_latents", None)
+        if reference_latents is not None:
+            out['reference_latent'] = comfy.conds.CONDRegular(self.process_latent_in(reference_latents[-1])[:, :, 0])
+
+        return out
+
+
+class WAN21_Vace(WAN21):
+    def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
+        super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.VaceWanModel)
+        self.image_to_video = image_to_video
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        noise = kwargs.get("noise", None)
+        noise_shape = list(noise.shape)
+        vace_frames = kwargs.get("vace_frames", None)
+        if vace_frames is None:
+            noise_shape[1] = 32
+            vace_frames = [torch.zeros(noise_shape, device=noise.device, dtype=noise.dtype)]
+
+        mask = kwargs.get("vace_mask", None)
+        if mask is None:
+            noise_shape[1] = 64
+            mask = [torch.ones(noise_shape, device=noise.device, dtype=noise.dtype)] * len(vace_frames)
+
+        vace_frames_out = []
+        for j in range(len(vace_frames)):
+            vf = vace_frames[j].to(device=noise.device, dtype=noise.dtype, copy=True)
+            for i in range(0, vf.shape[1], 16):
+                vf[:, i:i + 16] = self.process_latent_in(vf[:, i:i + 16])
+            vf = torch.cat([vf, mask[j].to(device=noise.device, dtype=noise.dtype)], dim=1)
+            vace_frames_out.append(vf)
+
+        vace_frames = torch.stack(vace_frames_out, dim=1)
+        out['vace_context'] = comfy.conds.CONDRegular(vace_frames)
+
+        vace_strength = kwargs.get("vace_strength", [1.0] * len(vace_frames_out))
+        out['vace_strength'] = comfy.conds.CONDConstant(vace_strength)
+        return out
+
+class WAN21_Camera(WAN21):
+    def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
+        super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.CameraWanModel)
+        self.image_to_video = image_to_video
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        camera_conditions = kwargs.get("camera_conditions", None)
+        if camera_conditions is not None:
+            out['camera_conditions'] = comfy.conds.CONDRegular(camera_conditions)
+        return out
+
+class WAN22_S2V(WAN21):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel_S2V)
+        self.memory_usage_factor_conds = ("reference_latent", "reference_motion")
+        self.memory_usage_shape_process = {"reference_motion": lambda shape: [shape[0], shape[1], 1.5, shape[-2], shape[-1]]}
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        audio_embed = kwargs.get("audio_embed", None)
+        if audio_embed is not None:
+            out['audio_embed'] = comfy.conds.CONDRegular(audio_embed)
+
+        reference_latents = kwargs.get("reference_latents", None)
+        if reference_latents is not None:
+            out['reference_latent'] = comfy.conds.CONDRegular(self.process_latent_in(reference_latents[-1]))
+
+        reference_motion = kwargs.get("reference_motion", None)
+        if reference_motion is not None:
+            out['reference_motion'] = comfy.conds.CONDRegular(self.process_latent_in(reference_motion))
+
+        control_video = kwargs.get("control_video", None)
+        if control_video is not None:
+            out['control_video'] = comfy.conds.CONDRegular(self.process_latent_in(control_video))
+        return out
+
+    def extra_conds_shapes(self, **kwargs):
+        out = {}
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            out['reference_latent'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16])
+
+        reference_motion = kwargs.get("reference_motion", None)
+        if reference_motion is not None:
+            out['reference_motion'] = reference_motion.shape
+        return out
+
+class WAN22(WAN21):
+    def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
+        super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)
+        self.image_to_video = image_to_video
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        denoise_mask = kwargs.get("denoise_mask", None)
+        if denoise_mask is not None:
+            out["denoise_mask"] = comfy.conds.CONDRegular(denoise_mask)
+        return out
+
+    def process_timestep(self, timestep, x, denoise_mask=None, **kwargs):
+        if denoise_mask is None:
+            return timestep
+        temp_ts = (torch.mean(denoise_mask[:, :, :, :, :], dim=(1, 3, 4), keepdim=True) * timestep.view([timestep.shape[0]] + [1] * (denoise_mask.ndim - 1))).reshape(timestep.shape[0], -1)
+        return temp_ts
+
+    def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
+        return latent_image
+
+class Hunyuan3Dv2(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan3d.model.Hunyuan3Dv2)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        guidance = kwargs.get("guidance", 5.0)
+        if guidance is not None:
+            out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
+        return out
+
+class HiDream(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hidream.model.HiDreamImageTransformer2DModel)
+
+    def encode_adm(self, **kwargs):
+        return kwargs["pooled_output"]
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        conditioning_llama3 = kwargs.get("conditioning_llama3", None)
+        if conditioning_llama3 is not None:
+            out['encoder_hidden_states_llama3'] = comfy.conds.CONDRegular(conditioning_llama3)
+        image_cond = kwargs.get("concat_latent_image", None)
+        if image_cond is not None:
+            out['image_cond'] = comfy.conds.CONDNoiseShape(self.process_latent_in(image_cond))
+        return out
+
+class Chroma(Flux):
+    def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.chroma.model.Chroma)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+
+        guidance = kwargs.get("guidance", 0)
+        if guidance is not None:
+            out['guidance'] = comfy.conds.CONDRegular(torch.FloatTensor([guidance]))
+        return out
+
+class ACEStep(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.ace.model.ACEStepTransformer2DModel)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        noise = kwargs.get("noise", None)
+
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+
+        conditioning_lyrics = kwargs.get("conditioning_lyrics", None)
+        if cross_attn is not None:
+            out['lyric_token_idx'] = comfy.conds.CONDRegular(conditioning_lyrics)
+        out['speaker_embeds'] = comfy.conds.CONDRegular(torch.zeros(noise.shape[0], 512, device=noise.device, dtype=noise.dtype))
+        out['lyrics_strength'] = comfy.conds.CONDConstant(kwargs.get("lyrics_strength", 1.0))
+        return out
+
+class Omnigen2(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.omnigen.omnigen2.OmniGen2Transformer2DModel)
+        self.memory_usage_factor_conds = ("ref_latents",)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        attention_mask = kwargs.get("attention_mask", None)
+        if attention_mask is not None:
+            if torch.numel(attention_mask) != attention_mask.sum():
+                out['attention_mask'] = comfy.conds.CONDRegular(attention_mask)
+            out['num_tokens'] = comfy.conds.CONDConstant(max(1, torch.sum(attention_mask).item()))
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            latents = []
+            for lat in ref_latents:
+                latents.append(self.process_latent_in(lat))
+            out['ref_latents'] = comfy.conds.CONDList(latents)
+        return out
+
+    def extra_conds_shapes(self, **kwargs):
+        out = {}
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16])
+        return out
+
+class QwenImage(BaseModel):
+    def __init__(self, model_config, model_type=ModelType.FLUX, device=None):
+        super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.qwen_image.model.QwenImageTransformer2DModel)
+        self.memory_usage_factor_conds = ("ref_latents",)
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+        cross_attn = kwargs.get("cross_attn", None)
+        if cross_attn is not None:
+            out['c_crossattn'] = comfy.conds.CONDRegular(cross_attn)
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            latents = []
+            for lat in ref_latents:
+                latents.append(self.process_latent_in(lat))
+            out['ref_latents'] = comfy.conds.CONDList(latents)
+
+            ref_latents_method = kwargs.get("reference_latents_method", None)
+            if ref_latents_method is not None:
+                out['ref_latents_method'] = comfy.conds.CONDConstant(ref_latents_method)
+        return out
+
+    def extra_conds_shapes(self, **kwargs):
+        out = {}
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            out['ref_latents'] = list([1, 16, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16])
+        return out

comfy/model_detection.py

+import json
+import comfy.supported_models
+import comfy.supported_models_base
+import comfy.utils
+import math
+import logging
+import torch
+
+def count_blocks(state_dict_keys, prefix_string):
+    count = 0
+    while True:
+        c = False
+        for k in state_dict_keys:
+            if k.startswith(prefix_string.format(count)):
+                c = True
+                break
+        if c == False:
+            break
+        count += 1
+    return count
+
+def calculate_transformer_depth(prefix, state_dict_keys, state_dict):
+    context_dim = None
+    use_linear_in_transformer = False
+
+    transformer_prefix = prefix + "1.transformer_blocks."
+    transformer_keys = sorted(list(filter(lambda a: a.startswith(transformer_prefix), state_dict_keys)))
+    if len(transformer_keys) > 0:
+        last_transformer_depth = count_blocks(state_dict_keys, transformer_prefix + '{}')
+        context_dim = state_dict['{}0.attn2.to_k.weight'.format(transformer_prefix)].shape[1]
+        use_linear_in_transformer = len(state_dict['{}1.proj_in.weight'.format(prefix)].shape) == 2
+        time_stack = '{}1.time_stack.0.attn1.to_q.weight'.format(prefix) in state_dict or '{}1.time_mix_blocks.0.attn1.to_q.weight'.format(prefix) in state_dict
+        time_stack_cross = '{}1.time_stack.0.attn2.to_q.weight'.format(prefix) in state_dict or '{}1.time_mix_blocks.0.attn2.to_q.weight'.format(prefix) in state_dict
+        return last_transformer_depth, context_dim, use_linear_in_transformer, time_stack, time_stack_cross
+    return None
+
+def detect_unet_config(state_dict, key_prefix, metadata=None):
+    state_dict_keys = list(state_dict.keys())
+
+    if '{}joint_blocks.0.context_block.attn.qkv.weight'.format(key_prefix) in state_dict_keys: #mmdit model
+        unet_config = {}
+        unet_config["in_channels"] = state_dict['{}x_embedder.proj.weight'.format(key_prefix)].shape[1]
+        patch_size = state_dict['{}x_embedder.proj.weight'.format(key_prefix)].shape[2]
+        unet_config["patch_size"] = patch_size
+        final_layer = '{}final_layer.linear.weight'.format(key_prefix)
+        if final_layer in state_dict:
+            unet_config["out_channels"] = state_dict[final_layer].shape[0] // (patch_size * patch_size)
+
+        unet_config["depth"] = state_dict['{}x_embedder.proj.weight'.format(key_prefix)].shape[0] // 64
+        unet_config["input_size"] = None
+        y_key = '{}y_embedder.mlp.0.weight'.format(key_prefix)
+        if y_key in state_dict_keys:
+            unet_config["adm_in_channels"] = state_dict[y_key].shape[1]
+
+        context_key = '{}context_embedder.weight'.format(key_prefix)
+        if context_key in state_dict_keys:
+            in_features = state_dict[context_key].shape[1]
+            out_features = state_dict[context_key].shape[0]
+            unet_config["context_embedder_config"] = {"target": "torch.nn.Linear", "params": {"in_features": in_features, "out_features": out_features}}
+        num_patches_key = '{}pos_embed'.format(key_prefix)
+        if num_patches_key in state_dict_keys:
+            num_patches = state_dict[num_patches_key].shape[1]
+            unet_config["num_patches"] = num_patches
+            unet_config["pos_embed_max_size"] = round(math.sqrt(num_patches))
+
+        rms_qk = '{}joint_blocks.0.context_block.attn.ln_q.weight'.format(key_prefix)
+        if rms_qk in state_dict_keys:
+            unet_config["qk_norm"] = "rms"
+
+        unet_config["pos_embed_scaling_factor"] = None #unused for inference
+        context_processor = '{}context_processor.layers.0.attn.qkv.weight'.format(key_prefix)
+        if context_processor in state_dict_keys:
+            unet_config["context_processor_layers"] = count_blocks(state_dict_keys, '{}context_processor.layers.'.format(key_prefix) + '{}.')
+        unet_config["x_block_self_attn_layers"] = []
+        for key in state_dict_keys:
+            if key.startswith('{}joint_blocks.'.format(key_prefix)) and key.endswith('.x_block.attn2.qkv.weight'):
+                layer = key[len('{}joint_blocks.'.format(key_prefix)):-len('.x_block.attn2.qkv.weight')]
+                unet_config["x_block_self_attn_layers"].append(int(layer))
+        return unet_config
+
+    if '{}clf.1.weight'.format(key_prefix) in state_dict_keys: #stable cascade
+        unet_config = {}
+        text_mapper_name = '{}clip_txt_mapper.weight'.format(key_prefix)
+        if text_mapper_name in state_dict_keys:
+            unet_config['stable_cascade_stage'] = 'c'
+            w = state_dict[text_mapper_name]
+            if w.shape[0] == 1536: #stage c lite
+                unet_config['c_cond'] = 1536
+                unet_config['c_hidden'] = [1536, 1536]
+                unet_config['nhead'] = [24, 24]
+                unet_config['blocks'] = [[4, 12], [12, 4]]
+            elif w.shape[0] == 2048: #stage c full
+                unet_config['c_cond'] = 2048
+        elif '{}clip_mapper.weight'.format(key_prefix) in state_dict_keys:
+            unet_config['stable_cascade_stage'] = 'b'
+            w = state_dict['{}down_blocks.1.0.channelwise.0.weight'.format(key_prefix)]
+            if w.shape[-1] == 640:
+                unet_config['c_hidden'] = [320, 640, 1280, 1280]
+                unet_config['nhead'] = [-1, -1, 20, 20]
+                unet_config['blocks'] = [[2, 6, 28, 6], [6, 28, 6, 2]]
+                unet_config['block_repeat'] = [[1, 1, 1, 1], [3, 3, 2, 2]]
+            elif w.shape[-1] == 576: #stage b lite
+                unet_config['c_hidden'] = [320, 576, 1152, 1152]
+                unet_config['nhead'] = [-1, 9, 18, 18]
+                unet_config['blocks'] = [[2, 4, 14, 4], [4, 14, 4, 2]]
+                unet_config['block_repeat'] = [[1, 1, 1, 1], [2, 2, 2, 2]]
+        return unet_config
+
+    if '{}transformer.rotary_pos_emb.inv_freq'.format(key_prefix) in state_dict_keys: #stable audio dit
+        unet_config = {}
+        unet_config["audio_model"] = "dit1.0"
+        return unet_config
+
+    if '{}double_layers.0.attn.w1q.weight'.format(key_prefix) in state_dict_keys: #aura flow dit
+        unet_config = {}
+        unet_config["max_seq"] = state_dict['{}positional_encoding'.format(key_prefix)].shape[1]
+        unet_config["cond_seq_dim"] = state_dict['{}cond_seq_linear.weight'.format(key_prefix)].shape[1]
+        double_layers = count_blocks(state_dict_keys, '{}double_layers.'.format(key_prefix) + '{}.')
+        single_layers = count_blocks(state_dict_keys, '{}single_layers.'.format(key_prefix) + '{}.')
+        unet_config["n_double_layers"] = double_layers
+        unet_config["n_layers"] = double_layers + single_layers
+        return unet_config
+
+    if '{}mlp_t5.0.weight'.format(key_prefix) in state_dict_keys: #Hunyuan DiT
+        unet_config = {}
+        unet_config["image_model"] = "hydit"
+        unet_config["depth"] = count_blocks(state_dict_keys, '{}blocks.'.format(key_prefix) + '{}.')
+        unet_config["hidden_size"] = state_dict['{}x_embedder.proj.weight'.format(key_prefix)].shape[0]
+        if unet_config["hidden_size"] == 1408 and unet_config["depth"] == 40: #DiT-g/2
+            unet_config["mlp_ratio"] = 4.3637
+        if state_dict['{}extra_embedder.0.weight'.format(key_prefix)].shape[1] == 3968:
+            unet_config["size_cond"] = True
+            unet_config["use_style_cond"] = True
+            unet_config["image_model"] = "hydit1"
+        return unet_config
+
+    if '{}txt_in.individual_token_refiner.blocks.0.norm1.weight'.format(key_prefix) in state_dict_keys: #Hunyuan Video
+        dit_config = {}
+        dit_config["image_model"] = "hunyuan_video"
+        dit_config["in_channels"] = state_dict['{}img_in.proj.weight'.format(key_prefix)].shape[1] #SkyReels img2video has 32 input channels
+        dit_config["patch_size"] = [1, 2, 2]
+        dit_config["out_channels"] = 16
+        dit_config["vec_in_dim"] = 768
+        dit_config["context_in_dim"] = 4096
+        dit_config["hidden_size"] = 3072
+        dit_config["mlp_ratio"] = 4.0
+        dit_config["num_heads"] = 24
+        dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.')
+        dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.')
+        dit_config["axes_dim"] = [16, 56, 56]
+        dit_config["theta"] = 256
+        dit_config["qkv_bias"] = True
+        guidance_keys = list(filter(lambda a: a.startswith("{}guidance_in.".format(key_prefix)), state_dict_keys))
+        dit_config["guidance_embed"] = len(guidance_keys) > 0
+        return dit_config
+
+    if '{}double_blocks.0.img_attn.norm.key_norm.scale'.format(key_prefix) in state_dict_keys and '{}img_in.weight'.format(key_prefix) in state_dict_keys: #Flux
+        dit_config = {}
+        dit_config["image_model"] = "flux"
+        dit_config["in_channels"] = 16
+        patch_size = 2
+        dit_config["patch_size"] = patch_size
+        in_key = "{}img_in.weight".format(key_prefix)
+        if in_key in state_dict_keys:
+            dit_config["in_channels"] = state_dict[in_key].shape[1] // (patch_size * patch_size)
+        dit_config["out_channels"] = 16
+        vec_in_key = '{}vector_in.in_layer.weight'.format(key_prefix)
+        if vec_in_key in state_dict_keys:
+            dit_config["vec_in_dim"] = state_dict[vec_in_key].shape[1]
+        dit_config["context_in_dim"] = 4096
+        dit_config["hidden_size"] = 3072
+        dit_config["mlp_ratio"] = 4.0
+        dit_config["num_heads"] = 24
+        dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.')
+        dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.')
+        dit_config["axes_dim"] = [16, 56, 56]
+        dit_config["theta"] = 10000
+        dit_config["qkv_bias"] = True
+        if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma
+            dit_config["image_model"] = "chroma"
+            dit_config["in_channels"] = 64
+            dit_config["out_channels"] = 64
+            dit_config["in_dim"] = 64
+            dit_config["out_dim"] = 3072
+            dit_config["hidden_dim"] = 5120
+            dit_config["n_layers"] = 5
+        else:
+            dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
+        return dit_config
+
+    if '{}t5_yproj.weight'.format(key_prefix) in state_dict_keys: #Genmo mochi preview
+        dit_config = {}
+        dit_config["image_model"] = "mochi_preview"
+        dit_config["depth"] = 48
+        dit_config["patch_size"] = 2
+        dit_config["num_heads"] = 24
+        dit_config["hidden_size_x"] = 3072
+        dit_config["hidden_size_y"] = 1536
+        dit_config["mlp_ratio_x"] = 4.0
+        dit_config["mlp_ratio_y"] = 4.0
+        dit_config["learn_sigma"] = False
+        dit_config["in_channels"] = 12
+        dit_config["qk_norm"] = True
+        dit_config["qkv_bias"] = False
+        dit_config["out_bias"] = True
+        dit_config["attn_drop"] = 0.0
+        dit_config["patch_embed_bias"] = True
+        dit_config["posenc_preserve_area"] = True
+        dit_config["timestep_mlp_bias"] = True
+        dit_config["attend_to_padding"] = False
+        dit_config["timestep_scale"] = 1000.0
+        dit_config["use_t5"] = True
+        dit_config["t5_feat_dim"] = 4096
+        dit_config["t5_token_length"] = 256
+        dit_config["rope_theta"] = 10000.0
+        return dit_config
+
+    if '{}adaln_single.emb.timestep_embedder.linear_1.bias'.format(key_prefix) in state_dict_keys and '{}pos_embed.proj.bias'.format(key_prefix) in state_dict_keys:
+        # PixArt diffusers
+        return None
+
+    if '{}adaln_single.emb.timestep_embedder.linear_1.bias'.format(key_prefix) in state_dict_keys: #Lightricks ltxv
+        dit_config = {}
+        dit_config["image_model"] = "ltxv"
+        dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.')
+        shape = state_dict['{}transformer_blocks.0.attn2.to_k.weight'.format(key_prefix)].shape
+        dit_config["attention_head_dim"] = shape[0] // 32
+        dit_config["cross_attention_dim"] = shape[1]
+        if metadata is not None and "config" in metadata:
+            dit_config.update(json.loads(metadata["config"]).get("transformer", {}))
+        return dit_config
+
+    if '{}genre_embedder.weight'.format(key_prefix) in state_dict_keys: #ACE-Step model
+        dit_config = {}
+        dit_config["audio_model"] = "ace"
+        dit_config["attention_head_dim"] = 128
+        dit_config["in_channels"] = 8
+        dit_config["inner_dim"] = 2560
+        dit_config["max_height"] = 16
+        dit_config["max_position"] = 32768
+        dit_config["max_width"] = 32768
+        dit_config["mlp_ratio"] = 2.5
+        dit_config["num_attention_heads"] = 20
+        dit_config["num_layers"] = 24
+        dit_config["out_channels"] = 8
+        dit_config["patch_size"] = [16, 1]
+        dit_config["rope_theta"] = 1000000.0
+        dit_config["speaker_embedding_dim"] = 512
+        dit_config["text_embedding_dim"] = 768
+
+        dit_config["ssl_encoder_depths"] = [8, 8]
+        dit_config["ssl_latent_dims"] = [1024, 768]
+        dit_config["ssl_names"] = ["mert", "m-hubert"]
+        dit_config["lyric_encoder_vocab_size"] = 6693
+        dit_config["lyric_hidden_size"] = 1024
+        return dit_config
+
+    if '{}t_block.1.weight'.format(key_prefix) in state_dict_keys: # PixArt
+        patch_size = 2
+        dit_config = {}
+        dit_config["num_heads"] = 16
+        dit_config["patch_size"] = patch_size
+        dit_config["hidden_size"] = 1152
+        dit_config["in_channels"] = 4
+        dit_config["depth"] = count_blocks(state_dict_keys, '{}blocks.'.format(key_prefix) + '{}.')
+
+        y_key = "{}y_embedder.y_embedding".format(key_prefix)
+        if y_key in state_dict_keys:
+            dit_config["model_max_length"] = state_dict[y_key].shape[0]
+
+        pe_key = "{}pos_embed".format(key_prefix)
+        if pe_key in state_dict_keys:
+            dit_config["input_size"] = int(math.sqrt(state_dict[pe_key].shape[1])) * patch_size
+            dit_config["pe_interpolation"] = dit_config["input_size"] // (512//8) # guess
+
+        ar_key = "{}ar_embedder.mlp.0.weight".format(key_prefix)
+        if ar_key in state_dict_keys:
+            dit_config["image_model"] = "pixart_alpha"
+            dit_config["micro_condition"] = True
+        else:
+            dit_config["image_model"] = "pixart_sigma"
+            dit_config["micro_condition"] = False
+        return dit_config
+
+    if '{}blocks.block0.blocks.0.block.attn.to_q.0.weight'.format(key_prefix) in state_dict_keys:  # Cosmos
+        dit_config = {}
+        dit_config["image_model"] = "cosmos"
+        dit_config["max_img_h"] = 240
+        dit_config["max_img_w"] = 240
+        dit_config["max_frames"] = 128
+        concat_padding_mask = True
+        dit_config["in_channels"] = (state_dict['{}x_embedder.proj.1.weight'.format(key_prefix)].shape[1] // 4) - int(concat_padding_mask)
+        dit_config["out_channels"] = 16
+        dit_config["patch_spatial"] = 2
+        dit_config["patch_temporal"] = 1
+        dit_config["model_channels"] = state_dict['{}blocks.block0.blocks.0.block.attn.to_q.0.weight'.format(key_prefix)].shape[0]
+        dit_config["block_config"] = "FA-CA-MLP"
+        dit_config["concat_padding_mask"] = concat_padding_mask
+        dit_config["pos_emb_cls"] = "rope3d"
+        dit_config["pos_emb_learnable"] = False
+        dit_config["pos_emb_interpolation"] = "crop"
+        dit_config["block_x_format"] = "THWBD"
+        dit_config["affline_emb_norm"] = True
+        dit_config["use_adaln_lora"] = True
+        dit_config["adaln_lora_dim"] = 256
+
+        if dit_config["model_channels"] == 4096:
+            # 7B
+            dit_config["num_blocks"] = 28
+            dit_config["num_heads"] = 32
+            dit_config["extra_per_block_abs_pos_emb"] = True
+            dit_config["rope_h_extrapolation_ratio"] = 1.0
+            dit_config["rope_w_extrapolation_ratio"] = 1.0
+            dit_config["rope_t_extrapolation_ratio"] = 2.0
+            dit_config["extra_per_block_abs_pos_emb_type"] = "learnable"
+        else:  # 5120
+            # 14B
+            dit_config["num_blocks"] = 36
+            dit_config["num_heads"] = 40
+            dit_config["extra_per_block_abs_pos_emb"] = True
+            dit_config["rope_h_extrapolation_ratio"] = 2.0
+            dit_config["rope_w_extrapolation_ratio"] = 2.0
+            dit_config["rope_t_extrapolation_ratio"] = 2.0
+            dit_config["extra_h_extrapolation_ratio"] = 2.0
+            dit_config["extra_w_extrapolation_ratio"] = 2.0
+            dit_config["extra_t_extrapolation_ratio"] = 2.0
+            dit_config["extra_per_block_abs_pos_emb_type"] = "learnable"
+        return dit_config
+
+    if '{}cap_embedder.1.weight'.format(key_prefix) in state_dict_keys:  # Lumina 2
+        dit_config = {}
+        dit_config["image_model"] = "lumina2"
+        dit_config["patch_size"] = 2
+        dit_config["in_channels"] = 16
+        dit_config["dim"] = 2304
+        dit_config["cap_feat_dim"] = 2304
+        dit_config["n_layers"] = 26
+        dit_config["n_heads"] = 24
+        dit_config["n_kv_heads"] = 8
+        dit_config["qk_norm"] = True
+        dit_config["axes_dims"] = [32, 32, 32]
+        dit_config["axes_lens"] = [300, 512, 512]
+        return dit_config
+
+    if '{}head.modulation'.format(key_prefix) in state_dict_keys:  # Wan 2.1
+        dit_config = {}
+        dit_config["image_model"] = "wan2.1"
+        dim = state_dict['{}head.modulation'.format(key_prefix)].shape[-1]
+        out_dim = state_dict['{}head.head.weight'.format(key_prefix)].shape[0] // 4
+        dit_config["dim"] = dim
+        dit_config["out_dim"] = out_dim
+        dit_config["num_heads"] = dim // 128
+        dit_config["ffn_dim"] = state_dict['{}blocks.0.ffn.0.weight'.format(key_prefix)].shape[0]
+        dit_config["num_layers"] = count_blocks(state_dict_keys, '{}blocks.'.format(key_prefix) + '{}.')
+        dit_config["patch_size"] = (1, 2, 2)
+        dit_config["freq_dim"] = 256
+        dit_config["window_size"] = (-1, -1)
+        dit_config["qk_norm"] = True
+        dit_config["cross_attn_norm"] = True
+        dit_config["eps"] = 1e-6
+        dit_config["in_dim"] = state_dict['{}patch_embedding.weight'.format(key_prefix)].shape[1]
+        if '{}vace_patch_embedding.weight'.format(key_prefix) in state_dict_keys:
+            dit_config["model_type"] = "vace"
+            dit_config["vace_in_dim"] = state_dict['{}vace_patch_embedding.weight'.format(key_prefix)].shape[1]
+            dit_config["vace_layers"] = count_blocks(state_dict_keys, '{}vace_blocks.'.format(key_prefix) + '{}.')
+        elif '{}control_adapter.conv.weight'.format(key_prefix) in state_dict_keys:
+            if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys:
+                dit_config["model_type"] = "camera"
+            else:
+                dit_config["model_type"] = "camera_2.2"
+        elif '{}casual_audio_encoder.encoder.final_linear.weight'.format(key_prefix) in state_dict_keys:
+            dit_config["model_type"] = "s2v"
+        else:
+            if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys:
+                dit_config["model_type"] = "i2v"
+            else:
+                dit_config["model_type"] = "t2v"
+        flf_weight = state_dict.get('{}img_emb.emb_pos'.format(key_prefix))
+        if flf_weight is not None:
+            dit_config["flf_pos_embed_token_number"] = flf_weight.shape[1]
+
+        ref_conv_weight = state_dict.get('{}ref_conv.weight'.format(key_prefix))
+        if ref_conv_weight is not None:
+            dit_config["in_dim_ref_conv"] = ref_conv_weight.shape[1]
+
+        return dit_config
+
+    if '{}latent_in.weight'.format(key_prefix) in state_dict_keys:  # Hunyuan 3D
+        in_shape = state_dict['{}latent_in.weight'.format(key_prefix)].shape
+        dit_config = {}
+        dit_config["image_model"] = "hunyuan3d2"
+        dit_config["in_channels"] = in_shape[1]
+        dit_config["context_in_dim"] = state_dict['{}cond_in.weight'.format(key_prefix)].shape[1]
+        dit_config["hidden_size"] = in_shape[0]
+        dit_config["mlp_ratio"] = 4.0
+        dit_config["num_heads"] = 16
+        dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.')
+        dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.')
+        dit_config["qkv_bias"] = True
+        dit_config["guidance_embed"] = "{}guidance_in.in_layer.weight".format(key_prefix) in state_dict_keys
+        return dit_config
+
+    if '{}caption_projection.0.linear.weight'.format(key_prefix) in state_dict_keys:  # HiDream
+        dit_config = {}
+        dit_config["image_model"] = "hidream"
+        dit_config["attention_head_dim"] = 128
+        dit_config["axes_dims_rope"] = [64, 32, 32]
+        dit_config["caption_channels"] = [4096, 4096]
+        dit_config["max_resolution"] = [128, 128]
+        dit_config["in_channels"] = 16
+        dit_config["llama_layers"] = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31]
+        dit_config["num_attention_heads"] = 20
+        dit_config["num_routed_experts"] = 4
+        dit_config["num_activated_experts"] = 2
+        dit_config["num_layers"] = 16
+        dit_config["num_single_layers"] = 32
+        dit_config["out_channels"] = 16
+        dit_config["patch_size"] = 2
+        dit_config["text_emb_dim"] = 2048
+        return dit_config
+
+    if '{}blocks.0.mlp.layer1.weight'.format(key_prefix) in state_dict_keys:  # Cosmos predict2
+        dit_config = {}
+        dit_config["image_model"] = "cosmos_predict2"
+        dit_config["max_img_h"] = 240
+        dit_config["max_img_w"] = 240
+        dit_config["max_frames"] = 128
+        concat_padding_mask = True
+        dit_config["in_channels"] = (state_dict['{}x_embedder.proj.1.weight'.format(key_prefix)].shape[1] // 4) - int(concat_padding_mask)
+        dit_config["out_channels"] = 16
+        dit_config["patch_spatial"] = 2
+        dit_config["patch_temporal"] = 1
+        dit_config["model_channels"] = state_dict['{}x_embedder.proj.1.weight'.format(key_prefix)].shape[0]
+        dit_config["concat_padding_mask"] = concat_padding_mask
+        dit_config["crossattn_emb_channels"] = 1024
+        dit_config["pos_emb_cls"] = "rope3d"
+        dit_config["pos_emb_learnable"] = True
+        dit_config["pos_emb_interpolation"] = "crop"
+        dit_config["min_fps"] = 1
+        dit_config["max_fps"] = 30
+
+        dit_config["use_adaln_lora"] = True
+        dit_config["adaln_lora_dim"] = 256
+        if dit_config["model_channels"] == 2048:
+            dit_config["num_blocks"] = 28
+            dit_config["num_heads"] = 16
+        elif dit_config["model_channels"] == 5120:
+            dit_config["num_blocks"] = 36
+            dit_config["num_heads"] = 40
+
+        if dit_config["in_channels"] == 16:
+            dit_config["extra_per_block_abs_pos_emb"] = False
+            dit_config["rope_h_extrapolation_ratio"] = 4.0
+            dit_config["rope_w_extrapolation_ratio"] = 4.0
+            dit_config["rope_t_extrapolation_ratio"] = 1.0
+        elif dit_config["in_channels"] == 17: # img to video
+            if dit_config["model_channels"] == 2048:
+                dit_config["extra_per_block_abs_pos_emb"] = False
+                dit_config["rope_h_extrapolation_ratio"] = 3.0
+                dit_config["rope_w_extrapolation_ratio"] = 3.0
+                dit_config["rope_t_extrapolation_ratio"] = 1.0
+            elif dit_config["model_channels"] == 5120:
+                dit_config["rope_h_extrapolation_ratio"] = 2.0
+                dit_config["rope_w_extrapolation_ratio"] = 2.0
+                dit_config["rope_t_extrapolation_ratio"] = 0.8333333333333334
+
+        dit_config["extra_h_extrapolation_ratio"] = 1.0
+        dit_config["extra_w_extrapolation_ratio"] = 1.0
+        dit_config["extra_t_extrapolation_ratio"] = 1.0
+        dit_config["rope_enable_fps_modulation"] = False
+
+        return dit_config
+
+    if '{}time_caption_embed.timestep_embedder.linear_1.bias'.format(key_prefix) in state_dict_keys:  # Omnigen2
+        dit_config = {}
+        dit_config["image_model"] = "omnigen2"
+        dit_config["axes_dim_rope"] = [40, 40, 40]
+        dit_config["axes_lens"] = [1024, 1664, 1664]
+        dit_config["ffn_dim_multiplier"] = None
+        dit_config["hidden_size"] = 2520
+        dit_config["in_channels"] = 16
+        dit_config["multiple_of"] = 256
+        dit_config["norm_eps"] = 1e-05
+        dit_config["num_attention_heads"] = 21
+        dit_config["num_kv_heads"] = 7
+        dit_config["num_layers"] = 32
+        dit_config["num_refiner_layers"] = 2
+        dit_config["out_channels"] = None
+        dit_config["patch_size"] = 2
+        dit_config["text_feat_dim"] = 2048
+        dit_config["timestep_scale"] = 1000.0
+        return dit_config
+
+    if '{}txt_norm.weight'.format(key_prefix) in state_dict_keys:  # Qwen Image
+        dit_config = {}
+        dit_config["image_model"] = "qwen_image"
+        dit_config["in_channels"] = state_dict['{}img_in.weight'.format(key_prefix)].shape[1]
+        dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.')
+        return dit_config
+
+    if '{}input_blocks.0.0.weight'.format(key_prefix) not in state_dict_keys:
+        return None
+
+    unet_config = {
+        "use_checkpoint": False,
+        "image_size": 32,
+        "use_spatial_transformer": True,
+        "legacy": False
+    }
+
+    y_input = '{}label_emb.0.0.weight'.format(key_prefix)
+    if y_input in state_dict_keys:
+        unet_config["num_classes"] = "sequential"
+        unet_config["adm_in_channels"] = state_dict[y_input].shape[1]
+    else:
+        unet_config["adm_in_channels"] = None
+
+    model_channels = state_dict['{}input_blocks.0.0.weight'.format(key_prefix)].shape[0]
+    in_channels = state_dict['{}input_blocks.0.0.weight'.format(key_prefix)].shape[1]
+
+    out_key = '{}out.2.weight'.format(key_prefix)
+    if out_key in state_dict:
+        out_channels = state_dict[out_key].shape[0]
+    else:
+        out_channels = 4
+
+    num_res_blocks = []
+    channel_mult = []
+    transformer_depth = []
+    transformer_depth_output = []
+    context_dim = None
+    use_linear_in_transformer = False
+
+    video_model = False
+    video_model_cross = False
+
+    current_res = 1
+    count = 0
+
+    last_res_blocks = 0
+    last_channel_mult = 0
+
+    input_block_count = count_blocks(state_dict_keys, '{}input_blocks'.format(key_prefix) + '.{}.')
+    for count in range(input_block_count):
+        prefix = '{}input_blocks.{}.'.format(key_prefix, count)
+        prefix_output = '{}output_blocks.{}.'.format(key_prefix, input_block_count - count - 1)
+
+        block_keys = sorted(list(filter(lambda a: a.startswith(prefix), state_dict_keys)))
+        if len(block_keys) == 0:
+            break
+
+        block_keys_output = sorted(list(filter(lambda a: a.startswith(prefix_output), state_dict_keys)))
+
+        if "{}0.op.weight".format(prefix) in block_keys: #new layer
+            num_res_blocks.append(last_res_blocks)
+            channel_mult.append(last_channel_mult)
+
+            current_res *= 2
+            last_res_blocks = 0
+            last_channel_mult = 0
+            out = calculate_transformer_depth(prefix_output, state_dict_keys, state_dict)
+            if out is not None:
+                transformer_depth_output.append(out[0])
+            else:
+                transformer_depth_output.append(0)
+        else:
+            res_block_prefix = "{}0.in_layers.0.weight".format(prefix)
+            if res_block_prefix in block_keys:
+                last_res_blocks += 1
+                last_channel_mult = state_dict["{}0.out_layers.3.weight".format(prefix)].shape[0] // model_channels
+
+                out = calculate_transformer_depth(prefix, state_dict_keys, state_dict)
+                if out is not None:
+                    transformer_depth.append(out[0])
+                    if context_dim is None:
+                        context_dim = out[1]
+                        use_linear_in_transformer = out[2]
+                        video_model = out[3]
+                        video_model_cross = out[4]
+                else:
+                    transformer_depth.append(0)
+
+            res_block_prefix = "{}0.in_layers.0.weight".format(prefix_output)
+            if res_block_prefix in block_keys_output:
+                out = calculate_transformer_depth(prefix_output, state_dict_keys, state_dict)
+                if out is not None:
+                    transformer_depth_output.append(out[0])
+                else:
+                    transformer_depth_output.append(0)
+
+
+    num_res_blocks.append(last_res_blocks)
+    channel_mult.append(last_channel_mult)
+    if "{}middle_block.1.proj_in.weight".format(key_prefix) in state_dict_keys:
+        transformer_depth_middle = count_blocks(state_dict_keys, '{}middle_block.1.transformer_blocks.'.format(key_prefix) + '{}')
+    elif "{}middle_block.0.in_layers.0.weight".format(key_prefix) in state_dict_keys:
+        transformer_depth_middle = -1
+    else:
+        transformer_depth_middle = -2
+
+    unet_config["in_channels"] = in_channels
+    unet_config["out_channels"] = out_channels
+    unet_config["model_channels"] = model_channels
+    unet_config["num_res_blocks"] = num_res_blocks
+    unet_config["transformer_depth"] = transformer_depth
+    unet_config["transformer_depth_output"] = transformer_depth_output
+    unet_config["channel_mult"] = channel_mult
+    unet_config["transformer_depth_middle"] = transformer_depth_middle
+    unet_config['use_linear_in_transformer'] = use_linear_in_transformer
+    unet_config["context_dim"] = context_dim
+
+    if video_model:
+        unet_config["extra_ff_mix_layer"] = True
+        unet_config["use_spatial_context"] = True
+        unet_config["merge_strategy"] = "learned_with_images"
+        unet_config["merge_factor"] = 0.0
+        unet_config["video_kernel_size"] = [3, 1, 1]
+        unet_config["use_temporal_resblock"] = True
+        unet_config["use_temporal_attention"] = True
+        unet_config["disable_temporal_crossattention"] = not video_model_cross
+    else:
+        unet_config["use_temporal_resblock"] = False
+        unet_config["use_temporal_attention"] = False
+
+    return unet_config
+
+def model_config_from_unet_config(unet_config, state_dict=None):
+    for model_config in comfy.supported_models.models:
+        if model_config.matches(unet_config, state_dict):
+            return model_config(unet_config)
+
+    logging.error("no match {}".format(unet_config))
+    return None
+
+def model_config_from_unet(state_dict, unet_key_prefix, use_base_if_no_match=False, metadata=None):
+    unet_config = detect_unet_config(state_dict, unet_key_prefix, metadata=metadata)
+    if unet_config is None:
+        return None
+    model_config = model_config_from_unet_config(unet_config, state_dict)
+    if model_config is None and use_base_if_no_match:
+        model_config = comfy.supported_models_base.BASE(unet_config)
+
+    scaled_fp8_key = "{}scaled_fp8".format(unet_key_prefix)
+    if scaled_fp8_key in state_dict:
+        scaled_fp8_weight = state_dict.pop(scaled_fp8_key)
+        model_config.scaled_fp8 = scaled_fp8_weight.dtype
+        if model_config.scaled_fp8 == torch.float32:
+            model_config.scaled_fp8 = torch.float8_e4m3fn
+        if scaled_fp8_weight.nelement() == 2:
+            model_config.optimizations["fp8"] = False
+        else:
+            model_config.optimizations["fp8"] = True
+
+    return model_config
+
+def unet_prefix_from_state_dict(state_dict):
+    candidates = ["model.diffusion_model.", #ldm/sgm models
+                  "model.model.", #audio models
+                  "net.", #cosmos
+                  ]
+    counts = {k: 0 for k in candidates}
+    for k in state_dict:
+        for c in candidates:
+            if k.startswith(c):
+                counts[c] += 1
+                break
+
+    top = max(counts, key=counts.get)
+    if counts[top] > 5:
+        return top
+    else:
+        return "model." #aura flow and others
+
+
+def convert_config(unet_config):
+    new_config = unet_config.copy()
+    num_res_blocks = new_config.get("num_res_blocks", None)
+    channel_mult = new_config.get("channel_mult", None)
+
+    if isinstance(num_res_blocks, int):
+        num_res_blocks = len(channel_mult) * [num_res_blocks]
+
+    if "attention_resolutions" in new_config:
+        attention_resolutions = new_config.pop("attention_resolutions")
+        transformer_depth = new_config.get("transformer_depth", None)
+        transformer_depth_middle = new_config.get("transformer_depth_middle", None)
+
+        if isinstance(transformer_depth, int):
+            transformer_depth = len(channel_mult) * [transformer_depth]
+        if transformer_depth_middle is None:
+            transformer_depth_middle =  transformer_depth[-1]
+        t_in = []
+        t_out = []
+        s = 1
+        for i in range(len(num_res_blocks)):
+            res = num_res_blocks[i]
+            d = 0
+            if s in attention_resolutions:
+                d = transformer_depth[i]
+
+            t_in += [d] * res
+            t_out += [d] * (res + 1)
+            s *= 2
+        transformer_depth = t_in
+        new_config["transformer_depth"] = t_in
+        new_config["transformer_depth_output"] = t_out
+        new_config["transformer_depth_middle"] = transformer_depth_middle
+
+    new_config["num_res_blocks"] = num_res_blocks
+    return new_config
+
+
+def unet_config_from_diffusers_unet(state_dict, dtype=None):
+    if "conv_in.weight" not in state_dict:
+        return None
+
+    match = {}
+    transformer_depth = []
+
+    attn_res = 1
+    down_blocks = count_blocks(state_dict, "down_blocks.{}")
+    for i in range(down_blocks):
+        attn_blocks = count_blocks(state_dict, "down_blocks.{}.attentions.".format(i) + '{}')
+        res_blocks = count_blocks(state_dict, "down_blocks.{}.resnets.".format(i) + '{}')
+        for ab in range(attn_blocks):
+            transformer_count = count_blocks(state_dict, "down_blocks.{}.attentions.{}.transformer_blocks.".format(i, ab) + '{}')
+            transformer_depth.append(transformer_count)
+            if transformer_count > 0:
+                match["context_dim"] = state_dict["down_blocks.{}.attentions.{}.transformer_blocks.0.attn2.to_k.weight".format(i, ab)].shape[1]
+
+        attn_res *= 2
+        if attn_blocks == 0:
+            for i in range(res_blocks):
+                transformer_depth.append(0)
+
+    match["transformer_depth"] = transformer_depth
+
+    match["model_channels"] = state_dict["conv_in.weight"].shape[0]
+    match["in_channels"] = state_dict["conv_in.weight"].shape[1]
+    match["adm_in_channels"] = None
+    if "class_embedding.linear_1.weight" in state_dict:
+        match["adm_in_channels"] = state_dict["class_embedding.linear_1.weight"].shape[1]
+    elif "add_embedding.linear_1.weight" in state_dict:
+        match["adm_in_channels"] = state_dict["add_embedding.linear_1.weight"].shape[1]
+
+    SDXL = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+            'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
+            'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 10, 10], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 10,
+            'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 2, 2, 2, 10, 10, 10],
+            'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SDXL_refiner = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+                    'num_classes': 'sequential', 'adm_in_channels': 2560, 'dtype': dtype, 'in_channels': 4, 'model_channels': 384,
+                    'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [0, 0, 4, 4, 4, 4, 0, 0], 'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 4,
+                    'use_linear_in_transformer': True, 'context_dim': 1280, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 4, 4, 4, 4, 4, 4, 0, 0, 0],
+                    'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SD21 = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+            'adm_in_channels': None, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': [2, 2, 2, 2],
+            'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0], 'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1, 'use_linear_in_transformer': True,
+            'context_dim': 1024, 'num_head_channels': 64, 'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
+            'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SD21_uncliph = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+                    'num_classes': 'sequential', 'adm_in_channels': 2048, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
+                    'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0], 'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1,
+                    'use_linear_in_transformer': True, 'context_dim': 1024, 'num_head_channels': 64, 'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
+                    'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SD21_unclipl = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+                    'num_classes': 'sequential', 'adm_in_channels': 1536, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
+                    'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0], 'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1,
+                    'use_linear_in_transformer': True, 'context_dim': 1024, 'num_head_channels': 64, 'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
+                    'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SD15 = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 'adm_in_channels': None,
+            'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0],
+            'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1, 'use_linear_in_transformer': False, 'context_dim': 768, 'num_heads': 8,
+            'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
+            'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SDXL_mid_cnet = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+                     'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
+                     'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 0, 0, 1, 1], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 1,
+                     'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 0, 0, 0, 1, 1, 1],
+                     'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SDXL_small_cnet = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+                       'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
+                       'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 0, 0, 0, 0], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 0,
+                       'use_linear_in_transformer': True, 'num_head_channels': 64, 'context_dim': 1, 'transformer_depth_output': [0, 0, 0, 0, 0, 0, 0, 0, 0],
+                       'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SDXL_diffusers_inpaint = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+                              'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 9, 'model_channels': 320,
+                              'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 10, 10], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 10,
+                              'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 2, 2, 2, 10, 10, 10],
+                              'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SDXL_diffusers_ip2p = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+                              'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 8, 'model_channels': 320,
+                              'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 10, 10], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': 10,
+                              'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 0, 2, 2, 2, 10, 10, 10],
+                              'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SSD_1B = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+              'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
+              'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 2, 2, 4, 4], 'transformer_depth_output': [0, 0, 0, 1, 1, 2, 10, 4, 4],
+              'channel_mult': [1, 2, 4], 'transformer_depth_middle': -1, 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64,
+              'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    Segmind_Vega = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+              'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
+              'num_res_blocks': [2, 2, 2], 'transformer_depth': [0, 0, 1, 1, 2, 2], 'transformer_depth_output': [0, 0, 0, 1, 1, 1, 2, 2, 2],
+              'channel_mult': [1, 2, 4], 'transformer_depth_middle': -1, 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64,
+              'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    KOALA_700M = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+              'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
+              'num_res_blocks': [1, 1, 1], 'transformer_depth': [0, 2, 5], 'transformer_depth_output': [0, 0, 2, 2, 5, 5],
+              'channel_mult': [1, 2, 4], 'transformer_depth_middle': -2, 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64,
+              'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    KOALA_1B = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+              'num_classes': 'sequential', 'adm_in_channels': 2816, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320,
+              'num_res_blocks': [1, 1, 1], 'transformer_depth': [0, 2, 6], 'transformer_depth_output': [0, 0, 2, 2, 6, 6],
+              'channel_mult': [1, 2, 4], 'transformer_depth_middle': 6, 'use_linear_in_transformer': True, 'context_dim': 2048, 'num_head_channels': 64,
+              'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SD09_XS = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+            'adm_in_channels': None, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': [1, 1, 1],
+            'transformer_depth': [1, 1, 1], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': -2, 'use_linear_in_transformer': True,
+            'context_dim': 1024, 'num_head_channels': 64, 'transformer_depth_output': [1, 1, 1, 1, 1, 1],
+            'use_temporal_attention': False, 'use_temporal_resblock': False, 'disable_self_attentions': [True, False, False]}
+
+    SD_XS = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False,
+            'adm_in_channels': None, 'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': [1, 1, 1],
+            'transformer_depth': [0, 1, 1], 'channel_mult': [1, 2, 4], 'transformer_depth_middle': -2, 'use_linear_in_transformer': False,
+            'context_dim': 768, 'num_head_channels': 64, 'transformer_depth_output': [0, 0, 1, 1, 1, 1],
+            'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    SD15_diffusers_inpaint = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 'adm_in_channels': None,
+            'dtype': dtype, 'in_channels': 9, 'model_channels': 320, 'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0],
+            'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1, 'use_linear_in_transformer': False, 'context_dim': 768, 'num_heads': 8,
+            'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
+            'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    LotusD = {'use_checkpoint': False, 'image_size': 32, 'out_channels': 4, 'use_spatial_transformer': True, 'legacy': False, 'adm_in_channels': 4,
+            'dtype': dtype, 'in_channels': 4, 'model_channels': 320, 'num_res_blocks': [2, 2, 2, 2], 'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0],
+            'channel_mult': [1, 2, 4, 4], 'transformer_depth_middle': 1, 'use_linear_in_transformer': True, 'context_dim': 1024, 'num_heads': 8,
+            'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
+            'use_temporal_attention': False, 'use_temporal_resblock': False}
+
+    supported_models = [LotusD, SDXL, SDXL_refiner, SD21, SD15, SD21_uncliph, SD21_unclipl, SDXL_mid_cnet, SDXL_small_cnet, SDXL_diffusers_inpaint, SSD_1B, Segmind_Vega, KOALA_700M, KOALA_1B, SD09_XS, SD_XS, SDXL_diffusers_ip2p, SD15_diffusers_inpaint]
+
+    for unet_config in supported_models:
+        matches = True
+        for k in match:
+            if match[k] != unet_config[k]:
+                matches = False
+                break
+        if matches:
+            return convert_config(unet_config)
+    return None
+
+def model_config_from_diffusers_unet(state_dict):
+    unet_config = unet_config_from_diffusers_unet(state_dict)
+    if unet_config is not None:
+        return model_config_from_unet_config(unet_config)
+    return None
+
+def convert_diffusers_mmdit(state_dict, output_prefix=""):
+    out_sd = {}
+
+    if 'joint_transformer_blocks.0.attn.add_k_proj.weight' in state_dict: #AuraFlow
+        num_joint = count_blocks(state_dict, 'joint_transformer_blocks.{}.')
+        num_single = count_blocks(state_dict, 'single_transformer_blocks.{}.')
+        sd_map = comfy.utils.auraflow_to_diffusers({"n_double_layers": num_joint, "n_layers": num_joint + num_single}, output_prefix=output_prefix)
+    elif 'adaln_single.emb.timestep_embedder.linear_1.bias' in state_dict and 'pos_embed.proj.bias' in state_dict: # PixArt
+        num_blocks = count_blocks(state_dict, 'transformer_blocks.{}.')
+        sd_map = comfy.utils.pixart_to_diffusers({"depth": num_blocks}, output_prefix=output_prefix)
+    elif 'x_embedder.weight' in state_dict: #Flux
+        depth = count_blocks(state_dict, 'transformer_blocks.{}.')
+        depth_single_blocks = count_blocks(state_dict, 'single_transformer_blocks.{}.')
+        hidden_size = state_dict["x_embedder.bias"].shape[0]
+        sd_map = comfy.utils.flux_to_diffusers({"depth": depth, "depth_single_blocks": depth_single_blocks, "hidden_size": hidden_size}, output_prefix=output_prefix)
+    elif 'transformer_blocks.0.attn.add_q_proj.weight' in state_dict and 'pos_embed.proj.weight' in state_dict: #SD3
+        num_blocks = count_blocks(state_dict, 'transformer_blocks.{}.')
+        depth = state_dict["pos_embed.proj.weight"].shape[0] // 64
+        sd_map = comfy.utils.mmdit_to_diffusers({"depth": depth, "num_blocks": num_blocks}, output_prefix=output_prefix)
+    else:
+        return None
+
+    for k in sd_map:
+        weight = state_dict.get(k, None)
+        if weight is not None:
+            t = sd_map[k]
+
+            if not isinstance(t, str):
+                if len(t) > 2:
+                    fun = t[2]
+                else:
+                    fun = lambda a: a
+                offset = t[1]
+                if offset is not None:
+                    old_weight = out_sd.get(t[0], None)
+                    if old_weight is None:
+                        old_weight = torch.empty_like(weight)
+                    if old_weight.shape[offset[0]] < offset[1] + offset[2]:
+                        exp = list(weight.shape)
+                        exp[offset[0]] = offset[1] + offset[2]
+                        new = torch.empty(exp, device=weight.device, dtype=weight.dtype)
+                        new[:old_weight.shape[0]] = old_weight
+                        old_weight = new
+
+                    w = old_weight.narrow(offset[0], offset[1], offset[2])
+                else:
+                    old_weight = weight
+                    w = weight
+                w[:] = fun(weight)
+                t = t[0]
+                out_sd[t] = old_weight
+            else:
+                out_sd[t] = weight
+            state_dict.pop(k)
+
+    return out_sd