Initial commit

comfy/cldm/cldm.py

+#taken from: https://github.com/lllyasviel/ControlNet
+#and modified
+
+import torch
+import torch.nn as nn
+
+from ..ldm.modules.diffusionmodules.util import (
+    timestep_embedding,
+)
+
+from ..ldm.modules.attention import SpatialTransformer
+from ..ldm.modules.diffusionmodules.openaimodel import UNetModel, TimestepEmbedSequential, ResBlock, Downsample
+from ..ldm.util import exists
+from .control_types import UNION_CONTROLNET_TYPES
+from collections import OrderedDict
+import comfy.ops
+from comfy.ldm.modules.attention import optimized_attention
+
+class OptimizedAttention(nn.Module):
+    def __init__(self, c, nhead, dropout=0.0, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.heads = nhead
+        self.c = c
+
+        self.in_proj = operations.Linear(c, c * 3, bias=True, dtype=dtype, device=device)
+        self.out_proj = operations.Linear(c, c, bias=True, dtype=dtype, device=device)
+
+    def forward(self, x):
+        x = self.in_proj(x)
+        q, k, v = x.split(self.c, dim=2)
+        out = optimized_attention(q, k, v, self.heads)
+        return self.out_proj(out)
+
+class QuickGELU(nn.Module):
+    def forward(self, x: torch.Tensor):
+        return x * torch.sigmoid(1.702 * x)
+
+class ResBlockUnionControlnet(nn.Module):
+    def __init__(self, dim, nhead, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.attn = OptimizedAttention(dim, nhead, dtype=dtype, device=device, operations=operations)
+        self.ln_1 = operations.LayerNorm(dim, dtype=dtype, device=device)
+        self.mlp = nn.Sequential(
+            OrderedDict([("c_fc", operations.Linear(dim, dim * 4, dtype=dtype, device=device)), ("gelu", QuickGELU()),
+                         ("c_proj", operations.Linear(dim * 4, dim, dtype=dtype, device=device))]))
+        self.ln_2 = operations.LayerNorm(dim, dtype=dtype, device=device)
+
+    def attention(self, x: torch.Tensor):
+        return self.attn(x)
+
+    def forward(self, x: torch.Tensor):
+        x = x + self.attention(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+class ControlledUnetModel(UNetModel):
+    #implemented in the ldm unet
+    pass
+
+class ControlNet(nn.Module):
+    def __init__(
+        self,
+        image_size,
+        in_channels,
+        model_channels,
+        hint_channels,
+        num_res_blocks,
+        dropout=0,
+        channel_mult=(1, 2, 4, 8),
+        conv_resample=True,
+        dims=2,
+        num_classes=None,
+        use_checkpoint=False,
+        dtype=torch.float32,
+        num_heads=-1,
+        num_head_channels=-1,
+        num_heads_upsample=-1,
+        use_scale_shift_norm=False,
+        resblock_updown=False,
+        use_new_attention_order=False,
+        use_spatial_transformer=False,    # custom transformer support
+        transformer_depth=1,              # custom transformer support
+        context_dim=None,                 # custom transformer support
+        n_embed=None,                     # custom support for prediction of discrete ids into codebook of first stage vq model
+        legacy=True,
+        disable_self_attentions=None,
+        num_attention_blocks=None,
+        disable_middle_self_attn=False,
+        use_linear_in_transformer=False,
+        adm_in_channels=None,
+        transformer_depth_middle=None,
+        transformer_depth_output=None,
+        attn_precision=None,
+        union_controlnet_num_control_type=None,
+        device=None,
+        operations=comfy.ops.disable_weight_init,
+        **kwargs,
+    ):
+        super().__init__()
+        assert use_spatial_transformer == True, "use_spatial_transformer has to be true"
+        if use_spatial_transformer:
+            assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
+
+        if context_dim is not None:
+            assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
+            # from omegaconf.listconfig import ListConfig
+            # if type(context_dim) == ListConfig:
+            #     context_dim = list(context_dim)
+
+        if num_heads_upsample == -1:
+            num_heads_upsample = num_heads
+
+        if num_heads == -1:
+            assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
+
+        if num_head_channels == -1:
+            assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
+
+        self.dims = dims
+        self.image_size = image_size
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+
+        if isinstance(num_res_blocks, int):
+            self.num_res_blocks = len(channel_mult) * [num_res_blocks]
+        else:
+            if len(num_res_blocks) != len(channel_mult):
+                raise ValueError("provide num_res_blocks either as an int (globally constant) or "
+                                 "as a list/tuple (per-level) with the same length as channel_mult")
+            self.num_res_blocks = num_res_blocks
+
+        if disable_self_attentions is not None:
+            # should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
+            assert len(disable_self_attentions) == len(channel_mult)
+        if num_attention_blocks is not None:
+            assert len(num_attention_blocks) == len(self.num_res_blocks)
+            assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(len(num_attention_blocks))))
+
+        transformer_depth = transformer_depth[:]
+
+        self.dropout = dropout
+        self.channel_mult = channel_mult
+        self.conv_resample = conv_resample
+        self.num_classes = num_classes
+        self.use_checkpoint = use_checkpoint
+        self.dtype = dtype
+        self.num_heads = num_heads
+        self.num_head_channels = num_head_channels
+        self.num_heads_upsample = num_heads_upsample
+        self.predict_codebook_ids = n_embed is not None
+
+        time_embed_dim = model_channels * 4
+        self.time_embed = nn.Sequential(
+            operations.Linear(model_channels, time_embed_dim, dtype=self.dtype, device=device),
+            nn.SiLU(),
+            operations.Linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
+        )
+
+        if self.num_classes is not None:
+            if isinstance(self.num_classes, int):
+                self.label_emb = nn.Embedding(num_classes, time_embed_dim)
+            elif self.num_classes == "continuous":
+                self.label_emb = nn.Linear(1, time_embed_dim)
+            elif self.num_classes == "sequential":
+                assert adm_in_channels is not None
+                self.label_emb = nn.Sequential(
+                    nn.Sequential(
+                        operations.Linear(adm_in_channels, time_embed_dim, dtype=self.dtype, device=device),
+                        nn.SiLU(),
+                        operations.Linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
+                    )
+                )
+            else:
+                raise ValueError()
+
+        self.input_blocks = nn.ModuleList(
+            [
+                TimestepEmbedSequential(
+                    operations.conv_nd(dims, in_channels, model_channels, 3, padding=1, dtype=self.dtype, device=device)
+                )
+            ]
+        )
+        self.zero_convs = nn.ModuleList([self.make_zero_conv(model_channels, operations=operations, dtype=self.dtype, device=device)])
+
+        self.input_hint_block = TimestepEmbedSequential(
+                    operations.conv_nd(dims, hint_channels, 16, 3, padding=1, dtype=self.dtype, device=device),
+                    nn.SiLU(),
+                    operations.conv_nd(dims, 16, 16, 3, padding=1, dtype=self.dtype, device=device),
+                    nn.SiLU(),
+                    operations.conv_nd(dims, 16, 32, 3, padding=1, stride=2, dtype=self.dtype, device=device),
+                    nn.SiLU(),
+                    operations.conv_nd(dims, 32, 32, 3, padding=1, dtype=self.dtype, device=device),
+                    nn.SiLU(),
+                    operations.conv_nd(dims, 32, 96, 3, padding=1, stride=2, dtype=self.dtype, device=device),
+                    nn.SiLU(),
+                    operations.conv_nd(dims, 96, 96, 3, padding=1, dtype=self.dtype, device=device),
+                    nn.SiLU(),
+                    operations.conv_nd(dims, 96, 256, 3, padding=1, stride=2, dtype=self.dtype, device=device),
+                    nn.SiLU(),
+                    operations.conv_nd(dims, 256, model_channels, 3, padding=1, dtype=self.dtype, device=device)
+        )
+
+        self._feature_size = model_channels
+        input_block_chans = [model_channels]
+        ch = model_channels
+        ds = 1
+        for level, mult in enumerate(channel_mult):
+            for nr in range(self.num_res_blocks[level]):
+                layers = [
+                    ResBlock(
+                        ch,
+                        time_embed_dim,
+                        dropout,
+                        out_channels=mult * model_channels,
+                        dims=dims,
+                        use_checkpoint=use_checkpoint,
+                        use_scale_shift_norm=use_scale_shift_norm,
+                        dtype=self.dtype,
+                        device=device,
+                        operations=operations,
+                    )
+                ]
+                ch = mult * model_channels
+                num_transformers = transformer_depth.pop(0)
+                if num_transformers > 0:
+                    if num_head_channels == -1:
+                        dim_head = ch // num_heads
+                    else:
+                        num_heads = ch // num_head_channels
+                        dim_head = num_head_channels
+                    if legacy:
+                        #num_heads = 1
+                        dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+                    if exists(disable_self_attentions):
+                        disabled_sa = disable_self_attentions[level]
+                    else:
+                        disabled_sa = False
+
+                    if not exists(num_attention_blocks) or nr < num_attention_blocks[level]:
+                        layers.append(
+                            SpatialTransformer(
+                                ch, num_heads, dim_head, depth=num_transformers, context_dim=context_dim,
+                                disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
+                                use_checkpoint=use_checkpoint, attn_precision=attn_precision, dtype=self.dtype, device=device, operations=operations
+                            )
+                        )
+                self.input_blocks.append(TimestepEmbedSequential(*layers))
+                self.zero_convs.append(self.make_zero_conv(ch, operations=operations, dtype=self.dtype, device=device))
+                self._feature_size += ch
+                input_block_chans.append(ch)
+            if level != len(channel_mult) - 1:
+                out_ch = ch
+                self.input_blocks.append(
+                    TimestepEmbedSequential(
+                        ResBlock(
+                            ch,
+                            time_embed_dim,
+                            dropout,
+                            out_channels=out_ch,
+                            dims=dims,
+                            use_checkpoint=use_checkpoint,
+                            use_scale_shift_norm=use_scale_shift_norm,
+                            down=True,
+                            dtype=self.dtype,
+                            device=device,
+                            operations=operations
+                        )
+                        if resblock_updown
+                        else Downsample(
+                            ch, conv_resample, dims=dims, out_channels=out_ch, dtype=self.dtype, device=device, operations=operations
+                        )
+                    )
+                )
+                ch = out_ch
+                input_block_chans.append(ch)
+                self.zero_convs.append(self.make_zero_conv(ch, operations=operations, dtype=self.dtype, device=device))
+                ds *= 2
+                self._feature_size += ch
+
+        if num_head_channels == -1:
+            dim_head = ch // num_heads
+        else:
+            num_heads = ch // num_head_channels
+            dim_head = num_head_channels
+        if legacy:
+            #num_heads = 1
+            dim_head = ch // num_heads if use_spatial_transformer else num_head_channels
+        mid_block = [
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+                dtype=self.dtype,
+                device=device,
+                operations=operations
+            )]
+        if transformer_depth_middle >= 0:
+            mid_block += [SpatialTransformer(  # always uses a self-attn
+                            ch, num_heads, dim_head, depth=transformer_depth_middle, context_dim=context_dim,
+                            disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer,
+                            use_checkpoint=use_checkpoint, attn_precision=attn_precision, dtype=self.dtype, device=device, operations=operations
+                        ),
+            ResBlock(
+                ch,
+                time_embed_dim,
+                dropout,
+                dims=dims,
+                use_checkpoint=use_checkpoint,
+                use_scale_shift_norm=use_scale_shift_norm,
+                dtype=self.dtype,
+                device=device,
+                operations=operations
+            )]
+        self.middle_block = TimestepEmbedSequential(*mid_block)
+        self.middle_block_out = self.make_zero_conv(ch, operations=operations, dtype=self.dtype, device=device)
+        self._feature_size += ch
+
+        if union_controlnet_num_control_type is not None:
+            self.num_control_type = union_controlnet_num_control_type
+            num_trans_channel = 320
+            num_trans_head = 8
+            num_trans_layer = 1
+            num_proj_channel = 320
+            # task_scale_factor = num_trans_channel ** 0.5
+            self.task_embedding = nn.Parameter(torch.empty(self.num_control_type, num_trans_channel, dtype=self.dtype, device=device))
+
+            self.transformer_layes = nn.Sequential(*[ResBlockUnionControlnet(num_trans_channel, num_trans_head, dtype=self.dtype, device=device, operations=operations) for _ in range(num_trans_layer)])
+            self.spatial_ch_projs = operations.Linear(num_trans_channel, num_proj_channel, dtype=self.dtype, device=device)
+            #-----------------------------------------------------------------------------------------------------
+
+            control_add_embed_dim = 256
+            class ControlAddEmbedding(nn.Module):
+                def __init__(self, in_dim, out_dim, num_control_type, dtype=None, device=None, operations=None):
+                    super().__init__()
+                    self.num_control_type = num_control_type
+                    self.in_dim = in_dim
+                    self.linear_1 = operations.Linear(in_dim * num_control_type, out_dim, dtype=dtype, device=device)
+                    self.linear_2 = operations.Linear(out_dim, out_dim, dtype=dtype, device=device)
+                def forward(self, control_type, dtype, device):
+                    c_type = torch.zeros((self.num_control_type,), device=device)
+                    c_type[control_type] = 1.0
+                    c_type = timestep_embedding(c_type.flatten(), self.in_dim, repeat_only=False).to(dtype).reshape((-1, self.num_control_type * self.in_dim))
+                    return self.linear_2(torch.nn.functional.silu(self.linear_1(c_type)))
+
+            self.control_add_embedding = ControlAddEmbedding(control_add_embed_dim, time_embed_dim, self.num_control_type, dtype=self.dtype, device=device, operations=operations)
+        else:
+            self.task_embedding = None
+            self.control_add_embedding = None
+
+    def union_controlnet_merge(self, hint, control_type, emb, context):
+        # Equivalent to: https://github.com/xinsir6/ControlNetPlus/tree/main
+        inputs = []
+        condition_list = []
+
+        for idx in range(min(1, len(control_type))):
+            controlnet_cond = self.input_hint_block(hint[idx], emb, context)
+            feat_seq = torch.mean(controlnet_cond, dim=(2, 3))
+            if idx < len(control_type):
+                feat_seq += self.task_embedding[control_type[idx]].to(dtype=feat_seq.dtype, device=feat_seq.device)
+
+            inputs.append(feat_seq.unsqueeze(1))
+            condition_list.append(controlnet_cond)
+
+        x = torch.cat(inputs, dim=1)
+        x = self.transformer_layes(x)
+        controlnet_cond_fuser = None
+        for idx in range(len(control_type)):
+            alpha = self.spatial_ch_projs(x[:, idx])
+            alpha = alpha.unsqueeze(-1).unsqueeze(-1)
+            o = condition_list[idx] + alpha
+            if controlnet_cond_fuser is None:
+                controlnet_cond_fuser = o
+            else:
+                controlnet_cond_fuser += o
+        return controlnet_cond_fuser
+
+    def make_zero_conv(self, channels, operations=None, dtype=None, device=None):
+        return TimestepEmbedSequential(operations.conv_nd(self.dims, channels, channels, 1, padding=0, dtype=dtype, device=device))
+
+    def forward(self, x, hint, timesteps, context, y=None, **kwargs):
+        t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(x.dtype)
+        emb = self.time_embed(t_emb)
+
+        guided_hint = None
+        if self.control_add_embedding is not None: #Union Controlnet
+            control_type = kwargs.get("control_type", [])
+
+            if any([c >= self.num_control_type for c in control_type]):
+                max_type = max(control_type)
+                max_type_name = {
+                    v: k for k, v in UNION_CONTROLNET_TYPES.items()
+                }[max_type]
+                raise ValueError(
+                    f"Control type {max_type_name}({max_type}) is out of range for the number of control types" +
+                    f"({self.num_control_type}) supported.\n" +
+                    "Please consider using the ProMax ControlNet Union model.\n" +
+                    "https://huggingface.co/xinsir/controlnet-union-sdxl-1.0/tree/main"
+                )
+
+            emb += self.control_add_embedding(control_type, emb.dtype, emb.device)
+            if len(control_type) > 0:
+                if len(hint.shape) < 5:
+                    hint = hint.unsqueeze(dim=0)
+                guided_hint = self.union_controlnet_merge(hint, control_type, emb, context)
+
+        if guided_hint is None:
+            guided_hint = self.input_hint_block(hint, emb, context)
+
+        out_output = []
+        out_middle = []
+
+        if self.num_classes is not None:
+            assert y.shape[0] == x.shape[0]
+            emb = emb + self.label_emb(y)
+
+        h = x
+        for module, zero_conv in zip(self.input_blocks, self.zero_convs):
+            if guided_hint is not None:
+                h = module(h, emb, context)
+                h += guided_hint
+                guided_hint = None
+            else:
+                h = module(h, emb, context)
+            out_output.append(zero_conv(h, emb, context))
+
+        h = self.middle_block(h, emb, context)
+        out_middle.append(self.middle_block_out(h, emb, context))
+
+        return {"middle": out_middle, "output": out_output}
+

comfy/cldm/control_types.py

+UNION_CONTROLNET_TYPES = {
+    "openpose": 0,
+    "depth": 1,
+    "hed/pidi/scribble/ted": 2,
+    "canny/lineart/anime_lineart/mlsd": 3,
+    "normal": 4,
+    "segment": 5,
+    "tile": 6,
+    "repaint": 7,
+}

comfy/cldm/dit_embedder.py

+import math
+from typing import List, Optional, Tuple
+
+import torch
+import torch.nn as nn
+from torch import Tensor
+
+from comfy.ldm.modules.diffusionmodules.mmdit import DismantledBlock, PatchEmbed, VectorEmbedder, TimestepEmbedder, get_2d_sincos_pos_embed_torch
+
+
+class ControlNetEmbedder(nn.Module):
+
+    def __init__(
+        self,
+        img_size: int,
+        patch_size: int,
+        in_chans: int,
+        attention_head_dim: int,
+        num_attention_heads: int,
+        adm_in_channels: int,
+        num_layers: int,
+        main_model_double: int,
+        double_y_emb: bool,
+        device: torch.device,
+        dtype: torch.dtype,
+        pos_embed_max_size: Optional[int] = None,
+        operations = None,
+    ):
+        super().__init__()
+        self.main_model_double = main_model_double
+        self.dtype = dtype
+        self.hidden_size = num_attention_heads * attention_head_dim
+        self.patch_size = patch_size
+        self.x_embedder = PatchEmbed(
+            img_size=img_size,
+            patch_size=patch_size,
+            in_chans=in_chans,
+            embed_dim=self.hidden_size,
+            strict_img_size=pos_embed_max_size is None,
+            device=device,
+            dtype=dtype,
+            operations=operations,
+        )
+
+        self.t_embedder = TimestepEmbedder(self.hidden_size, dtype=dtype, device=device, operations=operations)
+
+        self.double_y_emb = double_y_emb
+        if self.double_y_emb:
+            self.orig_y_embedder = VectorEmbedder(
+                adm_in_channels, self.hidden_size, dtype, device, operations=operations
+            )
+            self.y_embedder = VectorEmbedder(
+                self.hidden_size, self.hidden_size, dtype, device, operations=operations
+            )
+        else:
+            self.y_embedder = VectorEmbedder(
+                adm_in_channels, self.hidden_size, dtype, device, operations=operations
+            )
+
+        self.transformer_blocks = nn.ModuleList(
+            DismantledBlock(
+                hidden_size=self.hidden_size, num_heads=num_attention_heads, qkv_bias=True,
+                dtype=dtype, device=device, operations=operations
+            )
+            for _ in range(num_layers)
+        )
+
+        # self.use_y_embedder = pooled_projection_dim != self.time_text_embed.text_embedder.linear_1.in_features
+        # TODO double check this logic when 8b
+        self.use_y_embedder = True
+
+        self.controlnet_blocks = nn.ModuleList([])
+        for _ in range(len(self.transformer_blocks)):
+            controlnet_block = operations.Linear(self.hidden_size, self.hidden_size, dtype=dtype, device=device)
+            self.controlnet_blocks.append(controlnet_block)
+
+        self.pos_embed_input = PatchEmbed(
+            img_size=img_size,
+            patch_size=patch_size,
+            in_chans=in_chans,
+            embed_dim=self.hidden_size,
+            strict_img_size=False,
+            device=device,
+            dtype=dtype,
+            operations=operations,
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        timesteps: torch.Tensor,
+        y: Optional[torch.Tensor] = None,
+        context: Optional[torch.Tensor] = None,
+        hint = None,
+    ) -> Tuple[Tensor, List[Tensor]]:
+        x_shape = list(x.shape)
+        x = self.x_embedder(x)
+        if not self.double_y_emb:
+            h = (x_shape[-2] + 1) // self.patch_size
+            w = (x_shape[-1] + 1) // self.patch_size
+            x += get_2d_sincos_pos_embed_torch(self.hidden_size, w, h, device=x.device)
+        c = self.t_embedder(timesteps, dtype=x.dtype)
+        if y is not None and self.y_embedder is not None:
+            if self.double_y_emb:
+                y = self.orig_y_embedder(y)
+            y = self.y_embedder(y)
+            c = c + y
+
+        x = x + self.pos_embed_input(hint)
+
+        block_out = ()
+
+        repeat = math.ceil(self.main_model_double / len(self.transformer_blocks))
+        for i in range(len(self.transformer_blocks)):
+            out = self.transformer_blocks[i](x, c)
+            if not self.double_y_emb:
+                x = out
+            block_out += (self.controlnet_blocks[i](out),) * repeat
+
+        return {"output": block_out}

comfy/cldm/mmdit.py

+import torch
+from typing import Optional
+import comfy.ldm.modules.diffusionmodules.mmdit
+
+class ControlNet(comfy.ldm.modules.diffusionmodules.mmdit.MMDiT):
+    def __init__(
+        self,
+        num_blocks = None,
+        control_latent_channels = None,
+        dtype = None,
+        device = None,
+        operations = None,
+        **kwargs,
+    ):
+        super().__init__(dtype=dtype, device=device, operations=operations, final_layer=False, num_blocks=num_blocks, **kwargs)
+        # controlnet_blocks
+        self.controlnet_blocks = torch.nn.ModuleList([])
+        for _ in range(len(self.joint_blocks)):
+            self.controlnet_blocks.append(operations.Linear(self.hidden_size, self.hidden_size, device=device, dtype=dtype))
+
+        if control_latent_channels is None:
+            control_latent_channels = self.in_channels
+
+        self.pos_embed_input = comfy.ldm.modules.diffusionmodules.mmdit.PatchEmbed(
+            None,
+            self.patch_size,
+            control_latent_channels,
+            self.hidden_size,
+            bias=True,
+            strict_img_size=False,
+            dtype=dtype,
+            device=device,
+            operations=operations
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        timesteps: torch.Tensor,
+        y: Optional[torch.Tensor] = None,
+        context: Optional[torch.Tensor] = None,
+        hint = None,
+    ) -> torch.Tensor:
+
+        #weird sd3 controlnet specific stuff
+        y = torch.zeros_like(y)
+
+        if self.context_processor is not None:
+            context = self.context_processor(context)
+
+        hw = x.shape[-2:]
+        x = self.x_embedder(x) + self.cropped_pos_embed(hw, device=x.device).to(dtype=x.dtype, device=x.device)
+        x += self.pos_embed_input(hint)
+
+        c = self.t_embedder(timesteps, dtype=x.dtype)
+        if y is not None and self.y_embedder is not None:
+            y = self.y_embedder(y)
+            c = c + y
+
+        if context is not None:
+            context = self.context_embedder(context)
+
+        output = []
+
+        blocks = len(self.joint_blocks)
+        for i in range(blocks):
+            context, x = self.joint_blocks[i](
+                context,
+                x,
+                c=c,
+                use_checkpoint=self.use_checkpoint,
+            )
+
+            out = self.controlnet_blocks[i](x)
+            count = self.depth // blocks
+            if i == blocks - 1:
+                count -= 1
+            for j in range(count):
+                output.append(out)
+
+        return {"output": output}

comfy/cli_args.py

+import argparse
+import enum
+import os
+import comfy.options
+
+
+class EnumAction(argparse.Action):
+    """
+    Argparse action for handling Enums
+    """
+    def __init__(self, **kwargs):
+        # Pop off the type value
+        enum_type = kwargs.pop("type", None)
+
+        # Ensure an Enum subclass is provided
+        if enum_type is None:
+            raise ValueError("type must be assigned an Enum when using EnumAction")
+        if not issubclass(enum_type, enum.Enum):
+            raise TypeError("type must be an Enum when using EnumAction")
+
+        # Generate choices from the Enum
+        choices = tuple(e.value for e in enum_type)
+        kwargs.setdefault("choices", choices)
+        kwargs.setdefault("metavar", f"[{','.join(list(choices))}]")
+
+        super(EnumAction, self).__init__(**kwargs)
+
+        self._enum = enum_type
+
+    def __call__(self, parser, namespace, values, option_string=None):
+        # Convert value back into an Enum
+        value = self._enum(values)
+        setattr(namespace, self.dest, value)
+
+
+parser = argparse.ArgumentParser()
+
+parser.add_argument("--listen", type=str, default="127.0.0.1", metavar="IP", nargs="?", const="0.0.0.0,::", help="Specify the IP address to listen on (default: 127.0.0.1). You can give a list of ip addresses by separating them with a comma like: 127.2.2.2,127.3.3.3 If --listen is provided without an argument, it defaults to 0.0.0.0,:: (listens on all ipv4 and ipv6)")
+parser.add_argument("--port", type=int, default=8188, help="Set the listen port.")
+parser.add_argument("--tls-keyfile", type=str, help="Path to TLS (SSL) key file. Enables TLS, makes app accessible at https://... requires --tls-certfile to function")
+parser.add_argument("--tls-certfile", type=str, help="Path to TLS (SSL) certificate file. Enables TLS, makes app accessible at https://... requires --tls-keyfile to function")
+parser.add_argument("--enable-cors-header", type=str, default=None, metavar="ORIGIN", nargs="?", const="*", help="Enable CORS (Cross-Origin Resource Sharing) with optional origin or allow all with default '*'.")
+parser.add_argument("--max-upload-size", type=float, default=100, help="Set the maximum upload size in MB.")
+
+parser.add_argument("--base-directory", type=str, default=None, help="Set the ComfyUI base directory for models, custom_nodes, input, output, temp, and user directories.")
+parser.add_argument("--extra-model-paths-config", type=str, default=None, metavar="PATH", nargs='+', action='append', help="Load one or more extra_model_paths.yaml files.")
+parser.add_argument("--output-directory", type=str, default=None, help="Set the ComfyUI output directory. Overrides --base-directory.")
+parser.add_argument("--temp-directory", type=str, default=None, help="Set the ComfyUI temp directory (default is in the ComfyUI directory). Overrides --base-directory.")
+parser.add_argument("--input-directory", type=str, default=None, help="Set the ComfyUI input directory. Overrides --base-directory.")
+parser.add_argument("--auto-launch", action="store_true", help="Automatically launch ComfyUI in the default browser.")
+parser.add_argument("--disable-auto-launch", action="store_true", help="Disable auto launching the browser.")
+parser.add_argument("--cuda-device", type=int, default=None, metavar="DEVICE_ID", help="Set the id of the cuda device this instance will use. All other devices will not be visible.")
+parser.add_argument("--default-device", type=int, default=None, metavar="DEFAULT_DEVICE_ID", help="Set the id of the default device, all other devices will stay visible.")
+cm_group = parser.add_mutually_exclusive_group()
+cm_group.add_argument("--cuda-malloc", action="store_true", help="Enable cudaMallocAsync (enabled by default for torch 2.0 and up).")
+cm_group.add_argument("--disable-cuda-malloc", action="store_true", help="Disable cudaMallocAsync.")
+
+
+fp_group = parser.add_mutually_exclusive_group()
+fp_group.add_argument("--force-fp32", action="store_true", help="Force fp32 (If this makes your GPU work better please report it).")
+fp_group.add_argument("--force-fp16", action="store_true", help="Force fp16.")
+
+fpunet_group = parser.add_mutually_exclusive_group()
+fpunet_group.add_argument("--fp32-unet", action="store_true", help="Run the diffusion model in fp32.")
+fpunet_group.add_argument("--fp64-unet", action="store_true", help="Run the diffusion model in fp64.")
+fpunet_group.add_argument("--bf16-unet", action="store_true", help="Run the diffusion model in bf16.")
+fpunet_group.add_argument("--fp16-unet", action="store_true", help="Run the diffusion model in fp16")
+fpunet_group.add_argument("--fp8_e4m3fn-unet", action="store_true", help="Store unet weights in fp8_e4m3fn.")
+fpunet_group.add_argument("--fp8_e5m2-unet", action="store_true", help="Store unet weights in fp8_e5m2.")
+fpunet_group.add_argument("--fp8_e8m0fnu-unet", action="store_true", help="Store unet weights in fp8_e8m0fnu.")
+
+fpvae_group = parser.add_mutually_exclusive_group()
+fpvae_group.add_argument("--fp16-vae", action="store_true", help="Run the VAE in fp16, might cause black images.")
+fpvae_group.add_argument("--fp32-vae", action="store_true", help="Run the VAE in full precision fp32.")
+fpvae_group.add_argument("--bf16-vae", action="store_true", help="Run the VAE in bf16.")
+
+parser.add_argument("--cpu-vae", action="store_true", help="Run the VAE on the CPU.")
+
+fpte_group = parser.add_mutually_exclusive_group()
+fpte_group.add_argument("--fp8_e4m3fn-text-enc", action="store_true", help="Store text encoder weights in fp8 (e4m3fn variant).")
+fpte_group.add_argument("--fp8_e5m2-text-enc", action="store_true", help="Store text encoder weights in fp8 (e5m2 variant).")
+fpte_group.add_argument("--fp16-text-enc", action="store_true", help="Store text encoder weights in fp16.")
+fpte_group.add_argument("--fp32-text-enc", action="store_true", help="Store text encoder weights in fp32.")
+fpte_group.add_argument("--bf16-text-enc", action="store_true", help="Store text encoder weights in bf16.")
+
+parser.add_argument("--force-channels-last", action="store_true", help="Force channels last format when inferencing the models.")
+
+parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE", const=-1, help="Use torch-directml.")
+
+parser.add_argument("--oneapi-device-selector", type=str, default=None, metavar="SELECTOR_STRING", help="Sets the oneAPI device(s) this instance will use.")
+parser.add_argument("--disable-ipex-optimize", action="store_true", help="Disables ipex.optimize default when loading models with Intel's Extension for Pytorch.")
+parser.add_argument("--supports-fp8-compute", action="store_true", help="ComfyUI will act like if the device supports fp8 compute.")
+
+class LatentPreviewMethod(enum.Enum):
+    NoPreviews = "none"
+    Auto = "auto"
+    Latent2RGB = "latent2rgb"
+    TAESD = "taesd"
+
+parser.add_argument("--preview-method", type=LatentPreviewMethod, default=LatentPreviewMethod.NoPreviews, help="Default preview method for sampler nodes.", action=EnumAction)
+
+parser.add_argument("--preview-size", type=int, default=512, help="Sets the maximum preview size for sampler nodes.")
+
+cache_group = parser.add_mutually_exclusive_group()
+cache_group.add_argument("--cache-classic", action="store_true", help="Use the old style (aggressive) caching.")
+cache_group.add_argument("--cache-lru", type=int, default=0, help="Use LRU caching with a maximum of N node results cached. May use more RAM/VRAM.")
+cache_group.add_argument("--cache-none", action="store_true", help="Reduced RAM/VRAM usage at the expense of executing every node for each run.")
+
+attn_group = parser.add_mutually_exclusive_group()
+attn_group.add_argument("--use-split-cross-attention", action="store_true", help="Use the split cross attention optimization. Ignored when xformers is used.")
+attn_group.add_argument("--use-quad-cross-attention", action="store_true", help="Use the sub-quadratic cross attention optimization . Ignored when xformers is used.")
+attn_group.add_argument("--use-pytorch-cross-attention", action="store_true", help="Use the new pytorch 2.0 cross attention function.")
+attn_group.add_argument("--use-sage-attention", action="store_true", help="Use sage attention.")
+attn_group.add_argument("--use-flash-attention", action="store_true", help="Use FlashAttention.")
+
+parser.add_argument("--disable-xformers", action="store_true", help="Disable xformers.")
+
+upcast = parser.add_mutually_exclusive_group()
+upcast.add_argument("--force-upcast-attention", action="store_true", help="Force enable attention upcasting, please report if it fixes black images.")
+upcast.add_argument("--dont-upcast-attention", action="store_true", help="Disable all upcasting of attention. Should be unnecessary except for debugging.")
+
+
+vram_group = parser.add_mutually_exclusive_group()
+vram_group.add_argument("--gpu-only", action="store_true", help="Store and run everything (text encoders/CLIP models, etc... on the GPU).")
+vram_group.add_argument("--highvram", action="store_true", help="By default models will be unloaded to CPU memory after being used. This option keeps them in GPU memory.")
+vram_group.add_argument("--normalvram", action="store_true", help="Used to force normal vram use if lowvram gets automatically enabled.")
+vram_group.add_argument("--lowvram", action="store_true", help="Split the unet in parts to use less vram.")
+vram_group.add_argument("--novram", action="store_true", help="When lowvram isn't enough.")
+vram_group.add_argument("--cpu", action="store_true", help="To use the CPU for everything (slow).")
+
+parser.add_argument("--reserve-vram", type=float, default=None, help="Set the amount of vram in GB you want to reserve for use by your OS/other software. By default some amount is reserved depending on your OS.")
+
+parser.add_argument("--async-offload", action="store_true", help="Use async weight offloading.")
+
+parser.add_argument("--force-non-blocking", action="store_true", help="Force ComfyUI to use non-blocking operations for all applicable tensors. This may improve performance on some non-Nvidia systems but can cause issues with some workflows.")
+
+parser.add_argument("--default-hashing-function", type=str, choices=['md5', 'sha1', 'sha256', 'sha512'], default='sha256', help="Allows you to choose the hash function to use for duplicate filename / contents comparison. Default is sha256.")
+
+parser.add_argument("--disable-smart-memory", action="store_true", help="Force ComfyUI to agressively offload to regular ram instead of keeping models in vram when it can.")
+parser.add_argument("--deterministic", action="store_true", help="Make pytorch use slower deterministic algorithms when it can. Note that this might not make images deterministic in all cases.")
+
+class PerformanceFeature(enum.Enum):
+    Fp16Accumulation = "fp16_accumulation"
+    Fp8MatrixMultiplication = "fp8_matrix_mult"
+    CublasOps = "cublas_ops"
+
+parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: fp16_accumulation fp8_matrix_mult cublas_ops")
+
+parser.add_argument("--mmap-torch-files", action="store_true", help="Use mmap when loading ckpt/pt files.")
+parser.add_argument("--disable-mmap", action="store_true", help="Don't use mmap when loading safetensors.")
+
+parser.add_argument("--dont-print-server", action="store_true", help="Don't print server output.")
+parser.add_argument("--quick-test-for-ci", action="store_true", help="Quick test for CI.")
+parser.add_argument("--windows-standalone-build", action="store_true", help="Windows standalone build: Enable convenient things that most people using the standalone windows build will probably enjoy (like auto opening the page on startup).")
+
+parser.add_argument("--disable-metadata", action="store_true", help="Disable saving prompt metadata in files.")
+parser.add_argument("--disable-all-custom-nodes", action="store_true", help="Disable loading all custom nodes.")
+parser.add_argument("--whitelist-custom-nodes", type=str, nargs='+', default=[], help="Specify custom node folders to load even when --disable-all-custom-nodes is enabled.")
+parser.add_argument("--disable-api-nodes", action="store_true", help="Disable loading all api nodes.")
+
+parser.add_argument("--multi-user", action="store_true", help="Enables per-user storage.")
+
+parser.add_argument("--verbose", default='INFO', const='DEBUG', nargs="?", choices=['DEBUG', 'INFO', 'WARNING', 'ERROR', 'CRITICAL'], help='Set the logging level')
+parser.add_argument("--log-stdout", action="store_true", help="Send normal process output to stdout instead of stderr (default).")
+
+# The default built-in provider hosted under web/
+DEFAULT_VERSION_STRING = "comfyanonymous/ComfyUI@latest"
+
+parser.add_argument(
+    "--front-end-version",
+    type=str,
+    default=DEFAULT_VERSION_STRING,
+    help="""
+    Specifies the version of the frontend to be used. This command needs internet connectivity to query and
+    download available frontend implementations from GitHub releases.
+
+    The version string should be in the format of:
+    [repoOwner]/[repoName]@[version]
+    where version is one of: "latest" or a valid version number (e.g. "1.0.0")
+    """,
+)
+
+def is_valid_directory(path: str) -> str:
+    """Validate if the given path is a directory, and check permissions."""
+    if not os.path.exists(path):
+        raise argparse.ArgumentTypeError(f"The path '{path}' does not exist.")
+    if not os.path.isdir(path):
+        raise argparse.ArgumentTypeError(f"'{path}' is not a directory.")
+    if not os.access(path, os.R_OK):
+        raise argparse.ArgumentTypeError(f"You do not have read permissions for '{path}'.")
+    return path
+
+parser.add_argument(
+    "--front-end-root",
+    type=is_valid_directory,
+    default=None,
+    help="The local filesystem path to the directory where the frontend is located. Overrides --front-end-version.",
+)
+
+parser.add_argument("--user-directory", type=is_valid_directory, default=None, help="Set the ComfyUI user directory with an absolute path. Overrides --base-directory.")
+
+parser.add_argument("--enable-compress-response-body", action="store_true", help="Enable compressing response body.")
+
+parser.add_argument(
+    "--comfy-api-base",
+    type=str,
+    default="https://api.comfy.org",
+    help="Set the base URL for the ComfyUI API.  (default: https://api.comfy.org)",
+)
+
+database_default_path = os.path.abspath(
+    os.path.join(os.path.dirname(__file__), "..", "user", "comfyui.db")
+)
+parser.add_argument("--database-url", type=str, default=f"sqlite:///{database_default_path}", help="Specify the database URL, e.g. for an in-memory database you can use 'sqlite:///:memory:'.")
+
+# chenxiadd
+parser.add_argument(
+    "--jsondir",
+    type=str,
+    nargs="?",
+    default=None, 
+)
+
+
+if comfy.options.args_parsing:
+    args = parser.parse_args()
+else:
+    args = parser.parse_args([])
+
+if args.windows_standalone_build:
+    args.auto_launch = True
+
+if args.disable_auto_launch:
+    args.auto_launch = False
+
+if args.force_fp16:
+    args.fp16_unet = True
+
+
+# '--fast' is not provided, use an empty set
+if args.fast is None:
+    args.fast = set()
+# '--fast' is provided with an empty list, enable all optimizations
+elif args.fast == []:
+    args.fast = set(PerformanceFeature)
+# '--fast' is provided with a list of performance features, use that list
+else:
+    args.fast = set(args.fast)

comfy/clip_config_bigg.json

+{
+  "architectures": [
+    "CLIPTextModel"
+  ],
+  "attention_dropout": 0.0,
+  "bos_token_id": 0,
+  "dropout": 0.0,
+  "eos_token_id": 49407,
+  "hidden_act": "gelu",
+  "hidden_size": 1280,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 5120,
+  "layer_norm_eps": 1e-05,
+  "max_position_embeddings": 77,
+  "model_type": "clip_text_model",
+  "num_attention_heads": 20,
+  "num_hidden_layers": 32,
+  "pad_token_id": 1,
+  "projection_dim": 1280,
+  "torch_dtype": "float32",
+  "vocab_size": 49408
+}

comfy/clip_model.py

+import torch
+from comfy.ldm.modules.attention import optimized_attention_for_device
+import comfy.ops
+
+class CLIPAttention(torch.nn.Module):
+    def __init__(self, embed_dim, heads, dtype, device, operations):
+        super().__init__()
+
+        self.heads = heads
+        self.q_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)
+        self.k_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)
+        self.v_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)
+
+        self.out_proj = operations.Linear(embed_dim, embed_dim, bias=True, dtype=dtype, device=device)
+
+    def forward(self, x, mask=None, optimized_attention=None):
+        q = self.q_proj(x)
+        k = self.k_proj(x)
+        v = self.v_proj(x)
+
+        out = optimized_attention(q, k, v, self.heads, mask)
+        return self.out_proj(out)
+
+ACTIVATIONS = {"quick_gelu": lambda a: a * torch.sigmoid(1.702 * a),
+               "gelu": torch.nn.functional.gelu,
+               "gelu_pytorch_tanh": lambda a: torch.nn.functional.gelu(a, approximate="tanh"),
+}
+
+class CLIPMLP(torch.nn.Module):
+    def __init__(self, embed_dim, intermediate_size, activation, dtype, device, operations):
+        super().__init__()
+        self.fc1 = operations.Linear(embed_dim, intermediate_size, bias=True, dtype=dtype, device=device)
+        self.activation = ACTIVATIONS[activation]
+        self.fc2 = operations.Linear(intermediate_size, embed_dim, bias=True, dtype=dtype, device=device)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.activation(x)
+        x = self.fc2(x)
+        return x
+
+class CLIPLayer(torch.nn.Module):
+    def __init__(self, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations):
+        super().__init__()
+        self.layer_norm1 = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
+        self.self_attn = CLIPAttention(embed_dim, heads, dtype, device, operations)
+        self.layer_norm2 = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
+        self.mlp = CLIPMLP(embed_dim, intermediate_size, intermediate_activation, dtype, device, operations)
+
+    def forward(self, x, mask=None, optimized_attention=None):
+        x += self.self_attn(self.layer_norm1(x), mask, optimized_attention)
+        x += self.mlp(self.layer_norm2(x))
+        return x
+
+
+class CLIPEncoder(torch.nn.Module):
+    def __init__(self, num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations):
+        super().__init__()
+        self.layers = torch.nn.ModuleList([CLIPLayer(embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations) for i in range(num_layers)])
+
+    def forward(self, x, mask=None, intermediate_output=None):
+        optimized_attention = optimized_attention_for_device(x.device, mask=mask is not None, small_input=True)
+
+        if intermediate_output is not None:
+            if intermediate_output < 0:
+                intermediate_output = len(self.layers) + intermediate_output
+
+        intermediate = None
+        for i, l in enumerate(self.layers):
+            x = l(x, mask, optimized_attention)
+            if i == intermediate_output:
+                intermediate = x.clone()
+        return x, intermediate
+
+class CLIPEmbeddings(torch.nn.Module):
+    def __init__(self, embed_dim, vocab_size=49408, num_positions=77, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.token_embedding = operations.Embedding(vocab_size, embed_dim, dtype=dtype, device=device)
+        self.position_embedding = operations.Embedding(num_positions, embed_dim, dtype=dtype, device=device)
+
+    def forward(self, input_tokens, dtype=torch.float32):
+        return self.token_embedding(input_tokens, out_dtype=dtype) + comfy.ops.cast_to(self.position_embedding.weight, dtype=dtype, device=input_tokens.device)
+
+
+class CLIPTextModel_(torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        num_layers = config_dict["num_hidden_layers"]
+        embed_dim = config_dict["hidden_size"]
+        heads = config_dict["num_attention_heads"]
+        intermediate_size = config_dict["intermediate_size"]
+        intermediate_activation = config_dict["hidden_act"]
+        num_positions = config_dict["max_position_embeddings"]
+        self.eos_token_id = config_dict["eos_token_id"]
+
+        super().__init__()
+        self.embeddings = CLIPEmbeddings(embed_dim, num_positions=num_positions, dtype=dtype, device=device, operations=operations)
+        self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations)
+        self.final_layer_norm = operations.LayerNorm(embed_dim, dtype=dtype, device=device)
+
+    def forward(self, input_tokens=None, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=torch.float32, embeds_info=[]):
+        if embeds is not None:
+            x = embeds + comfy.ops.cast_to(self.embeddings.position_embedding.weight, dtype=dtype, device=embeds.device)
+        else:
+            x = self.embeddings(input_tokens, dtype=dtype)
+
+        mask = None
+        if attention_mask is not None:
+            mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1])
+            mask = mask.masked_fill(mask.to(torch.bool), -torch.finfo(x.dtype).max)
+
+        causal_mask = torch.full((x.shape[1], x.shape[1]), -torch.finfo(x.dtype).max, dtype=x.dtype, device=x.device).triu_(1)
+
+        if mask is not None:
+            mask += causal_mask
+        else:
+            mask = causal_mask
+
+        x, i = self.encoder(x, mask=mask, intermediate_output=intermediate_output)
+        x = self.final_layer_norm(x)
+        if i is not None and final_layer_norm_intermediate:
+            i = self.final_layer_norm(i)
+
+        if num_tokens is not None:
+            pooled_output = x[list(range(x.shape[0])), list(map(lambda a: a - 1, num_tokens))]
+        else:
+            pooled_output = x[torch.arange(x.shape[0], device=x.device), (torch.round(input_tokens).to(dtype=torch.int, device=x.device) == self.eos_token_id).int().argmax(dim=-1),]
+        return x, i, pooled_output
+
+class CLIPTextModel(torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        self.num_layers = config_dict["num_hidden_layers"]
+        self.text_model = CLIPTextModel_(config_dict, dtype, device, operations)
+        embed_dim = config_dict["hidden_size"]
+        self.text_projection = operations.Linear(embed_dim, embed_dim, bias=False, dtype=dtype, device=device)
+        self.dtype = dtype
+
+    def get_input_embeddings(self):
+        return self.text_model.embeddings.token_embedding
+
+    def set_input_embeddings(self, embeddings):
+        self.text_model.embeddings.token_embedding = embeddings
+
+    def forward(self, *args, **kwargs):
+        x = self.text_model(*args, **kwargs)
+        out = self.text_projection(x[2])
+        return (x[0], x[1], out, x[2])
+
+
+class CLIPVisionEmbeddings(torch.nn.Module):
+    def __init__(self, embed_dim, num_channels=3, patch_size=14, image_size=224, model_type="", dtype=None, device=None, operations=None):
+        super().__init__()
+
+        num_patches = (image_size // patch_size) ** 2
+        if model_type == "siglip_vision_model":
+            self.class_embedding = None
+            patch_bias = True
+        else:
+            num_patches = num_patches + 1
+            self.class_embedding = torch.nn.Parameter(torch.empty(embed_dim, dtype=dtype, device=device))
+            patch_bias = False
+
+        self.patch_embedding = operations.Conv2d(
+            in_channels=num_channels,
+            out_channels=embed_dim,
+            kernel_size=patch_size,
+            stride=patch_size,
+            bias=patch_bias,
+            dtype=dtype,
+            device=device
+        )
+
+        self.position_embedding = operations.Embedding(num_patches, embed_dim, dtype=dtype, device=device)
+
+    def forward(self, pixel_values):
+        embeds = self.patch_embedding(pixel_values).flatten(2).transpose(1, 2)
+        if self.class_embedding is not None:
+            embeds = torch.cat([comfy.ops.cast_to_input(self.class_embedding, embeds).expand(pixel_values.shape[0], 1, -1), embeds], dim=1)
+        return embeds + comfy.ops.cast_to_input(self.position_embedding.weight, embeds)
+
+
+class CLIPVision(torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        num_layers = config_dict["num_hidden_layers"]
+        embed_dim = config_dict["hidden_size"]
+        heads = config_dict["num_attention_heads"]
+        intermediate_size = config_dict["intermediate_size"]
+        intermediate_activation = config_dict["hidden_act"]
+        model_type = config_dict["model_type"]
+
+        self.embeddings = CLIPVisionEmbeddings(embed_dim, config_dict["num_channels"], config_dict["patch_size"], config_dict["image_size"], model_type=model_type, dtype=dtype, device=device, operations=operations)
+        if model_type == "siglip_vision_model":
+            self.pre_layrnorm = lambda a: a
+            self.output_layernorm = True
+        else:
+            self.pre_layrnorm = operations.LayerNorm(embed_dim)
+            self.output_layernorm = False
+        self.encoder = CLIPEncoder(num_layers, embed_dim, heads, intermediate_size, intermediate_activation, dtype, device, operations)
+        self.post_layernorm = operations.LayerNorm(embed_dim)
+
+    def forward(self, pixel_values, attention_mask=None, intermediate_output=None):
+        x = self.embeddings(pixel_values)
+        x = self.pre_layrnorm(x)
+        #TODO: attention_mask?
+        x, i = self.encoder(x, mask=None, intermediate_output=intermediate_output)
+        if self.output_layernorm:
+            x = self.post_layernorm(x)
+            pooled_output = x
+        else:
+            pooled_output = self.post_layernorm(x[:, 0, :])
+        return x, i, pooled_output
+
+class LlavaProjector(torch.nn.Module):
+    def __init__(self, in_dim, out_dim, dtype, device, operations):
+        super().__init__()
+        self.linear_1 = operations.Linear(in_dim, out_dim, bias=True, device=device, dtype=dtype)
+        self.linear_2 = operations.Linear(out_dim, out_dim, bias=True, device=device, dtype=dtype)
+
+    def forward(self, x):
+        return self.linear_2(torch.nn.functional.gelu(self.linear_1(x[:, 1:])))
+
+class CLIPVisionModelProjection(torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        self.vision_model = CLIPVision(config_dict, dtype, device, operations)
+        if "projection_dim" in config_dict:
+            self.visual_projection = operations.Linear(config_dict["hidden_size"], config_dict["projection_dim"], bias=False)
+        else:
+            self.visual_projection = lambda a: a
+
+        if "llava3" == config_dict.get("projector_type", None):
+            self.multi_modal_projector = LlavaProjector(config_dict["hidden_size"], 4096, dtype, device, operations)
+        else:
+            self.multi_modal_projector = None
+
+    def forward(self, *args, **kwargs):
+        x = self.vision_model(*args, **kwargs)
+        out = self.visual_projection(x[2])
+        projected = None
+        if self.multi_modal_projector is not None:
+            projected = self.multi_modal_projector(x[1])
+
+        return (x[0], x[1], out, projected)

comfy/clip_vision.py

+from .utils import load_torch_file, transformers_convert, state_dict_prefix_replace
+import os
+import torch
+import json
+import logging
+
+import comfy.ops
+import comfy.model_patcher
+import comfy.model_management
+import comfy.utils
+import comfy.clip_model
+import comfy.image_encoders.dino2
+
+class Output:
+    def __getitem__(self, key):
+        return getattr(self, key)
+    def __setitem__(self, key, item):
+        setattr(self, key, item)
+
+def clip_preprocess(image, size=224, mean=[0.48145466, 0.4578275, 0.40821073], std=[0.26862954, 0.26130258, 0.27577711], crop=True):
+    image = image[:, :, :, :3] if image.shape[3] > 3 else image
+    mean = torch.tensor(mean, device=image.device, dtype=image.dtype)
+    std = torch.tensor(std, device=image.device, dtype=image.dtype)
+    image = image.movedim(-1, 1)
+    if not (image.shape[2] == size and image.shape[3] == size):
+        if crop:
+            scale = (size / min(image.shape[2], image.shape[3]))
+            scale_size = (round(scale * image.shape[2]), round(scale * image.shape[3]))
+        else:
+            scale_size = (size, size)
+
+        image = torch.nn.functional.interpolate(image, size=scale_size, mode="bicubic", antialias=True)
+        h = (image.shape[2] - size)//2
+        w = (image.shape[3] - size)//2
+        image = image[:,:,h:h+size,w:w+size]
+    image = torch.clip((255. * image), 0, 255).round() / 255.0
+    return (image - mean.view([3,1,1])) / std.view([3,1,1])
+
+IMAGE_ENCODERS = {
+    "clip_vision_model": comfy.clip_model.CLIPVisionModelProjection,
+    "siglip_vision_model": comfy.clip_model.CLIPVisionModelProjection,
+    "dinov2": comfy.image_encoders.dino2.Dinov2Model,
+}
+
+class ClipVisionModel():
+    def __init__(self, json_config):
+        with open(json_config) as f:
+            config = json.load(f)
+
+        self.image_size = config.get("image_size", 224)
+        self.image_mean = config.get("image_mean", [0.48145466, 0.4578275, 0.40821073])
+        self.image_std = config.get("image_std", [0.26862954, 0.26130258, 0.27577711])
+        model_class = IMAGE_ENCODERS.get(config.get("model_type", "clip_vision_model"))
+        self.load_device = comfy.model_management.text_encoder_device()
+        offload_device = comfy.model_management.text_encoder_offload_device()
+        self.dtype = comfy.model_management.text_encoder_dtype(self.load_device)
+        self.model = model_class(config, self.dtype, offload_device, comfy.ops.manual_cast)
+        self.model.eval()
+
+        self.patcher = comfy.model_patcher.ModelPatcher(self.model, load_device=self.load_device, offload_device=offload_device)
+
+    def load_sd(self, sd):
+        return self.model.load_state_dict(sd, strict=False)
+
+    def get_sd(self):
+        return self.model.state_dict()
+
+    def encode_image(self, image, crop=True):
+        comfy.model_management.load_model_gpu(self.patcher)
+        pixel_values = clip_preprocess(image.to(self.load_device), size=self.image_size, mean=self.image_mean, std=self.image_std, crop=crop).float()
+        out = self.model(pixel_values=pixel_values, intermediate_output=-2)
+
+        outputs = Output()
+        outputs["last_hidden_state"] = out[0].to(comfy.model_management.intermediate_device())
+        outputs["image_embeds"] = out[2].to(comfy.model_management.intermediate_device())
+        outputs["penultimate_hidden_states"] = out[1].to(comfy.model_management.intermediate_device())
+        outputs["mm_projected"] = out[3]
+        return outputs
+
+def convert_to_transformers(sd, prefix):
+    sd_k = sd.keys()
+    if "{}transformer.resblocks.0.attn.in_proj_weight".format(prefix) in sd_k:
+        keys_to_replace = {
+            "{}class_embedding".format(prefix): "vision_model.embeddings.class_embedding",
+            "{}conv1.weight".format(prefix): "vision_model.embeddings.patch_embedding.weight",
+            "{}positional_embedding".format(prefix): "vision_model.embeddings.position_embedding.weight",
+            "{}ln_post.bias".format(prefix): "vision_model.post_layernorm.bias",
+            "{}ln_post.weight".format(prefix): "vision_model.post_layernorm.weight",
+            "{}ln_pre.bias".format(prefix): "vision_model.pre_layrnorm.bias",
+            "{}ln_pre.weight".format(prefix): "vision_model.pre_layrnorm.weight",
+        }
+
+        for x in keys_to_replace:
+            if x in sd_k:
+                sd[keys_to_replace[x]] = sd.pop(x)
+
+        if "{}proj".format(prefix) in sd_k:
+            sd['visual_projection.weight'] = sd.pop("{}proj".format(prefix)).transpose(0, 1)
+
+        sd = transformers_convert(sd, prefix, "vision_model.", 48)
+    else:
+        replace_prefix = {prefix: ""}
+        sd = state_dict_prefix_replace(sd, replace_prefix)
+    return sd
+
+def load_clipvision_from_sd(sd, prefix="", convert_keys=False):
+    if convert_keys:
+        sd = convert_to_transformers(sd, prefix)
+    if "vision_model.encoder.layers.47.layer_norm1.weight" in sd:
+        json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_g.json")
+    elif "vision_model.encoder.layers.30.layer_norm1.weight" in sd:
+        json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_h.json")
+    elif "vision_model.encoder.layers.22.layer_norm1.weight" in sd:
+        embed_shape = sd["vision_model.embeddings.position_embedding.weight"].shape[0]
+        if sd["vision_model.encoder.layers.0.layer_norm1.weight"].shape[0] == 1152:
+            if embed_shape == 729:
+                json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_384.json")
+            elif embed_shape == 1024:
+                json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_siglip_512.json")
+        elif embed_shape == 577:
+            if "multi_modal_projector.linear_1.bias" in sd:
+                json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl_336_llava.json")
+            else:
+                json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl_336.json")
+        else:
+            json_config = os.path.join(os.path.dirname(os.path.realpath(__file__)), "clip_vision_config_vitl.json")
+    elif "embeddings.patch_embeddings.projection.weight" in sd:
+        json_config = os.path.join(os.path.join(os.path.dirname(os.path.realpath(__file__)), "image_encoders"), "dino2_giant.json")
+    else:
+        return None
+
+    clip = ClipVisionModel(json_config)
+    m, u = clip.load_sd(sd)
+    if len(m) > 0:
+        logging.warning("missing clip vision: {}".format(m))
+    u = set(u)
+    keys = list(sd.keys())
+    for k in keys:
+        if k not in u:
+            sd.pop(k)
+    return clip
+
+def load(ckpt_path):
+    sd = load_torch_file(ckpt_path)
+    if "visual.transformer.resblocks.0.attn.in_proj_weight" in sd:
+        return load_clipvision_from_sd(sd, prefix="visual.", convert_keys=True)
+    else:
+        return load_clipvision_from_sd(sd)

comfy/clip_vision_config_g.json

+{
+  "attention_dropout": 0.0,
+  "dropout": 0.0,
+  "hidden_act": "gelu",
+  "hidden_size": 1664,
+  "image_size": 224,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 8192,
+  "layer_norm_eps": 1e-05,
+  "model_type": "clip_vision_model",
+  "num_attention_heads": 16,
+  "num_channels": 3,
+  "num_hidden_layers": 48,
+  "patch_size": 14,
+  "projection_dim": 1280,
+  "torch_dtype": "float32"
+}

comfy/clip_vision_config_h.json

+{
+  "attention_dropout": 0.0,
+  "dropout": 0.0,
+  "hidden_act": "gelu",
+  "hidden_size": 1280,
+  "image_size": 224,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 5120,
+  "layer_norm_eps": 1e-05,
+  "model_type": "clip_vision_model",
+  "num_attention_heads": 16,
+  "num_channels": 3,
+  "num_hidden_layers": 32,
+  "patch_size": 14,
+  "projection_dim": 1024,
+  "torch_dtype": "float32"
+}

comfy/clip_vision_config_vitl.json

+{
+  "attention_dropout": 0.0,
+  "dropout": 0.0,
+  "hidden_act": "quick_gelu",
+  "hidden_size": 1024,
+  "image_size": 224,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-05,
+  "model_type": "clip_vision_model",
+  "num_attention_heads": 16,
+  "num_channels": 3,
+  "num_hidden_layers": 24,
+  "patch_size": 14,
+  "projection_dim": 768,
+  "torch_dtype": "float32"
+}

comfy/clip_vision_config_vitl_336.json

+{
+  "attention_dropout": 0.0,
+  "dropout": 0.0,
+  "hidden_act": "quick_gelu",
+  "hidden_size": 1024,
+  "image_size": 336,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-5,
+  "model_type": "clip_vision_model",
+  "num_attention_heads": 16,
+  "num_channels": 3,
+  "num_hidden_layers": 24,
+  "patch_size": 14,
+  "projection_dim": 768,
+  "torch_dtype": "float32"
+}

comfy/clip_vision_config_vitl_336_llava.json

+{
+  "attention_dropout": 0.0,
+  "dropout": 0.0,
+  "hidden_act": "quick_gelu",
+  "hidden_size": 1024,
+  "image_size": 336,
+  "initializer_factor": 1.0,
+  "initializer_range": 0.02,
+  "intermediate_size": 4096,
+  "layer_norm_eps": 1e-5,
+  "model_type": "clip_vision_model",
+  "num_attention_heads": 16,
+  "num_channels": 3,
+  "num_hidden_layers": 24,
+  "patch_size": 14,
+  "projection_dim": 768,
+  "projector_type": "llava3",
+  "torch_dtype": "float32"
+}

comfy/clip_vision_siglip_384.json

+{
+  "num_channels": 3,
+  "hidden_act": "gelu_pytorch_tanh",
+  "hidden_size": 1152,
+  "image_size": 384,
+  "intermediate_size": 4304,
+  "model_type": "siglip_vision_model",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 27,
+  "patch_size": 14,
+  "image_mean": [0.5, 0.5, 0.5],
+  "image_std": [0.5, 0.5, 0.5]
+}

comfy/clip_vision_siglip_512.json

+{
+  "num_channels": 3,
+  "hidden_act": "gelu_pytorch_tanh",
+  "hidden_size": 1152,
+  "image_size": 512,
+  "intermediate_size": 4304,
+  "model_type": "siglip_vision_model",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 27,
+  "patch_size": 16,
+  "image_mean": [0.5, 0.5, 0.5],
+  "image_std": [0.5, 0.5, 0.5]
+}

comfy/comfy_types/README.md

+# Comfy Typing
+## Type hinting for ComfyUI Node development
+
+This module provides type hinting and concrete convenience types for node developers.
+If cloned to the custom_nodes directory of ComfyUI, types can be imported using:
+
+```python
+from comfy.comfy_types import IO, ComfyNodeABC, CheckLazyMixin
+
+class ExampleNode(ComfyNodeABC):
+    @classmethod
+    def INPUT_TYPES(s) -> InputTypeDict:
+        return {"required": {}}
+```
+
+Full example is in [examples/example_nodes.py](examples/example_nodes.py).
+
+# Types
+A few primary types are documented below.  More complete information is available via the docstrings on each type.
+
+## `IO`
+
+A string enum of built-in and a few custom data types.  Includes the following special types and their requisite plumbing:
+
+- `ANY`: `"*"`
+- `NUMBER`: `"FLOAT,INT"`
+- `PRIMITIVE`: `"STRING,FLOAT,INT,BOOLEAN"`
+
+## `ComfyNodeABC`
+
+An abstract base class for nodes, offering type-hinting / autocomplete, and somewhat-alright docstrings.
+
+### Type hinting for `INPUT_TYPES`
+
+![INPUT_TYPES auto-completion in Visual Studio Code](examples/input_types.png)
+
+### `INPUT_TYPES` return dict
+
+![INPUT_TYPES return value type hinting in Visual Studio Code](examples/required_hint.png)
+
+### Options for individual inputs
+
+![INPUT_TYPES return value option auto-completion in Visual Studio Code](examples/input_options.png)

comfy/comfy_types/__init__.py

+import torch
+from typing import Callable, Protocol, TypedDict, Optional, List
+from .node_typing import IO, InputTypeDict, ComfyNodeABC, CheckLazyMixin, FileLocator
+
+
+class UnetApplyFunction(Protocol):
+    """Function signature protocol on comfy.model_base.BaseModel.apply_model"""
+
+    def __call__(self, x: torch.Tensor, t: torch.Tensor, **kwargs) -> torch.Tensor:
+        pass
+
+
+class UnetApplyConds(TypedDict):
+    """Optional conditions for unet apply function."""
+
+    c_concat: Optional[torch.Tensor]
+    c_crossattn: Optional[torch.Tensor]
+    control: Optional[torch.Tensor]
+    transformer_options: Optional[dict]
+
+
+class UnetParams(TypedDict):
+    # Tensor of shape [B, C, H, W]
+    input: torch.Tensor
+    # Tensor of shape [B]
+    timestep: torch.Tensor
+    c: UnetApplyConds
+    # List of [0, 1], [0], [1], ...
+    # 0 means conditional, 1 means conditional unconditional
+    cond_or_uncond: List[int]
+
+
+UnetWrapperFunction = Callable[[UnetApplyFunction, UnetParams], torch.Tensor]
+
+
+__all__ = [
+    "UnetWrapperFunction",
+    UnetApplyConds.__name__,
+    UnetParams.__name__,
+    UnetApplyFunction.__name__,
+    IO.__name__,
+    InputTypeDict.__name__,
+    ComfyNodeABC.__name__,
+    CheckLazyMixin.__name__,
+    FileLocator.__name__,
+]

comfy/comfy_types/examples/example_nodes.py

+from comfy.comfy_types import IO, ComfyNodeABC, InputTypeDict
+from inspect import cleandoc
+
+
+class ExampleNode(ComfyNodeABC):
+    """An example node that just adds 1 to an input integer.
+
+    * Requires a modern IDE to provide any benefit (detail: an IDE configured with analysis paths etc).
+    * This node is intended as an example for developers only.
+    """
+
+    DESCRIPTION = cleandoc(__doc__)
+    CATEGORY = "examples"
+
+    @classmethod
+    def INPUT_TYPES(s) -> InputTypeDict:
+        return {
+            "required": {
+                "input_int": (IO.INT, {"defaultInput": True}),
+            }
+        }
+
+    RETURN_TYPES = (IO.INT,)
+    RETURN_NAMES = ("input_plus_one",)
+    FUNCTION = "execute"
+
+    def execute(self, input_int: int):
+        return (input_int + 1,)