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