Upload 388 files

ComfyUI/CODEOWNERS

@@ -0,0 +1 @@
+*       @comfyanonymous

ComfyUI/CONTRIBUTING.md

@@ -0,0 +1,41 @@
+# Contributing to ComfyUI
+
+Welcome, and thank you for your interest in contributing to ComfyUI!
+
+There are several ways in which you can contribute, beyond writing code. The goal of this document is to provide a high-level overview of how you can get involved.
+
+## Asking Questions
+
+Have a question? Instead of opening an issue, please ask on [Discord](https://comfy.org/discord) or [Matrix](https://app.element.io/#/room/%23comfyui_space%3Amatrix.org) channels. Our team and the community will help you.
+
+## Providing Feedback
+
+Your comments and feedback are welcome, and the development team is available via a handful of different channels.
+
+See the `#bug-report`, `#feature-request` and `#feedback` channels on Discord.
+
+## Reporting Issues
+
+Have you identified a reproducible problem in ComfyUI? Do you have a feature request? We want to hear about it! Here's how you can report your issue as effectively as possible.
+
+
+### Look For an Existing Issue
+
+Before you create a new issue, please do a search in [open issues](https://github.com/comfyanonymous/ComfyUI/issues) to see if the issue or feature request has already been filed.
+
+If you find your issue already exists, make relevant comments and add your [reaction](https://github.com/blog/2119-add-reactions-to-pull-requests-issues-and-comments). Use a reaction in place of a "+1" comment:
+
+* 👍 - upvote
+* 👎 - downvote
+
+If you cannot find an existing issue that describes your bug or feature, create a new issue. We have an issue template in place to organize new issues.
+
+
+### Creating Pull Requests
+
+* Please refer to the article on [creating pull requests](https://github.com/comfyanonymous/ComfyUI/wiki/How-to-Contribute-Code) and contributing to this project.
+
+
+## Thank You
+
+Your contributions to open source, large or small, make great projects like this possible. Thank you for taking the time to contribute.

ComfyUI/comfy/checkpoint_pickle.py

@@ -0,0 +1,13 @@
+import pickle
+
+load = pickle.load
+
+class Empty:
+    pass
+
+class Unpickler(pickle.Unpickler):
+    def find_class(self, module, name):
+        #TODO: safe unpickle
+        if module.startswith("pytorch_lightning"):
+            return Empty
+        return super().find_class(module, name)

ComfyUI/comfy/cldm/cldm.py

@@ -0,0 +1,437 @@
+#taken from: https://github.com/lllyasviel/ControlNet
+#and modified
+
+import torch
+import torch as th
+import torch.nn as nn
+
+from ..ldm.modules.diffusionmodules.util import (
+    zero_module,
+    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":
+                print("setting up linear c_adm embedding layer")
+                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 = []
+
+        hs = []
+        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}
+

ComfyUI/comfy/cldm/control_types.py

@@ -0,0 +1,10 @@
+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,
+}

ComfyUI/comfy/cldm/mmdit.py

@@ -0,0 +1,77 @@
+import torch
+from typing import Dict, Optional
+import comfy.ldm.modules.diffusionmodules.mmdit
+
+class ControlNet(comfy.ldm.modules.diffusionmodules.mmdit.MMDiT):
+    def __init__(
+        self,
+        num_blocks = 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))
+
+        self.pos_embed_input = comfy.ldm.modules.diffusionmodules.mmdit.PatchEmbed(
+            None,
+            self.patch_size,
+            self.in_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}

ComfyUI/comfy/cli_args.py

@@ -0,0 +1,180 @@
+import argparse
+import enum
+import os
+from typing import Optional
+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). If --listen is provided without an argument, it defaults to 0.0.0.0. (listens on all)")
+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("--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.")
+parser.add_argument("--temp-directory", type=str, default=None, help="Set the ComfyUI temp directory (default is in the ComfyUI directory).")
+parser.add_argument("--input-directory", type=str, default=None, help="Set the ComfyUI input 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.")
+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("--bf16-unet", action="store_true", help="Run the UNET in bf16. This should only be used for testing stuff.")
+fpunet_group.add_argument("--fp16-unet", action="store_true", help="Store unet weights 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.")
+
+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.")
+
+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("--disable-ipex-optimize", action="store_true", help="Disables ipex.optimize when loading models with Intel GPUs.")
+
+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)
+
+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.")
+
+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("--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.")
+
+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("--multi-user", action="store_true", help="Enables per-user storage.")
+
+parser.add_argument("--verbose", action="store_true", help="Enables more debug prints.")
+
+# 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: Optional[str]) -> Optional[str]:
+    """Validate if the given path is a directory."""
+    if path is None:
+        return None
+
+    if not os.path.isdir(path):
+        raise argparse.ArgumentTypeError(f"{path} is not a valid directory.")
+    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.",
+)
+
+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
+
+import logging
+logging_level = logging.INFO
+if args.verbose:
+    logging_level = logging.DEBUG
+
+logging.basicConfig(format="%(message)s", level=logging_level)

ComfyUI/comfy/clip_config_bigg.json

@@ -0,0 +1,23 @@
+{
+  "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
+}

ComfyUI/comfy/clip_model.py

@@ -0,0 +1,196 @@
+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,
+}
+
+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"]
+        self.eos_token_id = config_dict["eos_token_id"]
+
+        super().__init__()
+        self.embeddings = CLIPEmbeddings(embed_dim, 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, attention_mask=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=torch.float32):
+        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), float("-inf"))
+
+        causal_mask = torch.empty(x.shape[1], x.shape[1], dtype=x.dtype, device=x.device).fill_(float("-inf")).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)
+
+        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.text_projection.weight.copy_(torch.eye(embed_dim))
+        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, dtype=None, device=None, operations=None):
+        super().__init__()
+        self.class_embedding = torch.nn.Parameter(torch.empty(embed_dim, dtype=dtype, device=device))
+
+        self.patch_embedding = operations.Conv2d(
+            in_channels=num_channels,
+            out_channels=embed_dim,
+            kernel_size=patch_size,
+            stride=patch_size,
+            bias=False,
+            dtype=dtype,
+            device=device
+        )
+
+        num_patches = (image_size // patch_size) ** 2
+        num_positions = num_patches + 1
+        self.position_embedding = operations.Embedding(num_positions, embed_dim, dtype=dtype, device=device)
+
+    def forward(self, pixel_values):
+        embeds = self.patch_embedding(pixel_values).flatten(2).transpose(1, 2)
+        return torch.cat([comfy.ops.cast_to_input(self.class_embedding, embeds).expand(pixel_values.shape[0], 1, -1), embeds], dim=1) + 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"]
+
+        self.embeddings = CLIPVisionEmbeddings(embed_dim, config_dict["num_channels"], config_dict["patch_size"], config_dict["image_size"], dtype=dtype, device=device, operations=operations)
+        self.pre_layrnorm = operations.LayerNorm(embed_dim)
+        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)
+        pooled_output = self.post_layernorm(x[:, 0, :])
+        return x, i, pooled_output
+
+class CLIPVisionModelProjection(torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        self.vision_model = CLIPVision(config_dict, dtype, device, operations)
+        self.visual_projection = operations.Linear(config_dict["hidden_size"], config_dict["projection_dim"], bias=False)
+
+    def forward(self, *args, **kwargs):
+        x = self.vision_model(*args, **kwargs)
+        out = self.visual_projection(x[2])
+        return (x[0], x[1], out)

ComfyUI/comfy/clip_vision.py

@@ -0,0 +1,121 @@
+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
+
+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 = torch.tensor([ 0.48145466,0.4578275,0.40821073], device=image.device, dtype=image.dtype)
+    std = torch.tensor([0.26862954,0.26130258,0.27577711], device=image.device, dtype=image.dtype)
+    image = image.movedim(-1, 1)
+    if not (image.shape[2] == size and image.shape[3] == size):
+        scale = (size / min(image.shape[2], image.shape[3]))
+        image = torch.nn.functional.interpolate(image, size=(round(scale * image.shape[2]), round(scale * image.shape[3])), 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])
+
+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.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 = comfy.clip_model.CLIPVisionModelProjection(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):
+        comfy.model_management.load_model_gpu(self.patcher)
+        pixel_values = clip_preprocess(image.to(self.load_device), size=self.image_size).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())
+        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:
+        if sd["vision_model.embeddings.position_embedding.weight"].shape[0] == 577:
+            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")
+    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:
+            t = sd.pop(k)
+            del t
+    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)

ComfyUI/comfy/clip_vision_config_g.json

@@ -0,0 +1,18 @@
+{
+  "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"
+}

ComfyUI/comfy/clip_vision_config_h.json

@@ -0,0 +1,18 @@
+{
+  "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"
+}

ComfyUI/comfy/clip_vision_config_vitl.json

@@ -0,0 +1,18 @@
+{
+  "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"
+}

ComfyUI/comfy/clip_vision_config_vitl_336.json

@@ -0,0 +1,18 @@
+{
+  "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"
+}

ComfyUI/comfy/conds.py

@@ -0,0 +1,83 @@
+import torch
+import math
+import comfy.utils
+
+
+def lcm(a, b): #TODO: eventually replace by math.lcm (added in python3.9)
+    return abs(a*b) // math.gcd(a, b)
+
+class CONDRegular:
+    def __init__(self, cond):
+        self.cond = cond
+
+    def _copy_with(self, cond):
+        return self.__class__(cond)
+
+    def process_cond(self, batch_size, device, **kwargs):
+        return self._copy_with(comfy.utils.repeat_to_batch_size(self.cond, batch_size).to(device))
+
+    def can_concat(self, other):
+        if self.cond.shape != other.cond.shape:
+            return False
+        return True
+
+    def concat(self, others):
+        conds = [self.cond]
+        for x in others:
+            conds.append(x.cond)
+        return torch.cat(conds)
+
+class CONDNoiseShape(CONDRegular):
+    def process_cond(self, batch_size, device, area, **kwargs):
+        data = self.cond
+        if area is not None:
+            dims = len(area) // 2
+            for i in range(dims):
+                data = data.narrow(i + 2, area[i + dims], area[i])
+
+        return self._copy_with(comfy.utils.repeat_to_batch_size(data, batch_size).to(device))
+
+
+class CONDCrossAttn(CONDRegular):
+    def can_concat(self, other):
+        s1 = self.cond.shape
+        s2 = other.cond.shape
+        if s1 != s2:
+            if s1[0] != s2[0] or s1[2] != s2[2]: #these 2 cases should not happen
+                return False
+
+            mult_min = lcm(s1[1], s2[1])
+            diff = mult_min // min(s1[1], s2[1])
+            if diff > 4: #arbitrary limit on the padding because it's probably going to impact performance negatively if it's too much
+                return False
+        return True
+
+    def concat(self, others):
+        conds = [self.cond]
+        crossattn_max_len = self.cond.shape[1]
+        for x in others:
+            c = x.cond
+            crossattn_max_len = lcm(crossattn_max_len, c.shape[1])
+            conds.append(c)
+
+        out = []
+        for c in conds:
+            if c.shape[1] < crossattn_max_len:
+                c = c.repeat(1, crossattn_max_len // c.shape[1], 1) #padding with repeat doesn't change result
+            out.append(c)
+        return torch.cat(out)
+
+class CONDConstant(CONDRegular):
+    def __init__(self, cond):
+        self.cond = cond
+
+    def process_cond(self, batch_size, device, **kwargs):
+        return self._copy_with(self.cond)
+
+    def can_concat(self, other):
+        if self.cond != other.cond:
+            return False
+        return True
+
+    def concat(self, others):
+        return self.cond

ComfyUI/comfy/controlnet.py

@@ -0,0 +1,622 @@
+import torch
+import math
+import os
+import logging
+import comfy.utils
+import comfy.model_management
+import comfy.model_detection
+import comfy.model_patcher
+import comfy.ops
+import comfy.latent_formats
+
+import comfy.cldm.cldm
+import comfy.t2i_adapter.adapter
+import comfy.ldm.cascade.controlnet
+import comfy.cldm.mmdit
+
+
+def broadcast_image_to(tensor, target_batch_size, batched_number):
+    current_batch_size = tensor.shape[0]
+    #print(current_batch_size, target_batch_size)
+    if current_batch_size == 1:
+        return tensor
+
+    per_batch = target_batch_size // batched_number
+    tensor = tensor[:per_batch]
+
+    if per_batch > tensor.shape[0]:
+        tensor = torch.cat([tensor] * (per_batch // tensor.shape[0]) + [tensor[:(per_batch % tensor.shape[0])]], dim=0)
+
+    current_batch_size = tensor.shape[0]
+    if current_batch_size == target_batch_size:
+        return tensor
+    else:
+        return torch.cat([tensor] * batched_number, dim=0)
+
+class ControlBase:
+    def __init__(self, device=None):
+        self.cond_hint_original = None
+        self.cond_hint = None
+        self.strength = 1.0
+        self.timestep_percent_range = (0.0, 1.0)
+        self.latent_format = None
+        self.vae = None
+        self.global_average_pooling = False
+        self.timestep_range = None
+        self.compression_ratio = 8
+        self.upscale_algorithm = 'nearest-exact'
+        self.extra_args = {}
+
+        if device is None:
+            device = comfy.model_management.get_torch_device()
+        self.device = device
+        self.previous_controlnet = None
+
+    def set_cond_hint(self, cond_hint, strength=1.0, timestep_percent_range=(0.0, 1.0), vae=None):
+        self.cond_hint_original = cond_hint
+        self.strength = strength
+        self.timestep_percent_range = timestep_percent_range
+        if self.latent_format is not None:
+            self.vae = vae
+        return self
+
+    def pre_run(self, model, percent_to_timestep_function):
+        self.timestep_range = (percent_to_timestep_function(self.timestep_percent_range[0]), percent_to_timestep_function(self.timestep_percent_range[1]))
+        if self.previous_controlnet is not None:
+            self.previous_controlnet.pre_run(model, percent_to_timestep_function)
+
+    def set_previous_controlnet(self, controlnet):
+        self.previous_controlnet = controlnet
+        return self
+
+    def cleanup(self):
+        if self.previous_controlnet is not None:
+            self.previous_controlnet.cleanup()
+        if self.cond_hint is not None:
+            del self.cond_hint
+            self.cond_hint = None
+        self.timestep_range = None
+
+    def get_models(self):
+        out = []
+        if self.previous_controlnet is not None:
+            out += self.previous_controlnet.get_models()
+        return out
+
+    def copy_to(self, c):
+        c.cond_hint_original = self.cond_hint_original
+        c.strength = self.strength
+        c.timestep_percent_range = self.timestep_percent_range
+        c.global_average_pooling = self.global_average_pooling
+        c.compression_ratio = self.compression_ratio
+        c.upscale_algorithm = self.upscale_algorithm
+        c.latent_format = self.latent_format
+        c.extra_args = self.extra_args.copy()
+        c.vae = self.vae
+
+    def inference_memory_requirements(self, dtype):
+        if self.previous_controlnet is not None:
+            return self.previous_controlnet.inference_memory_requirements(dtype)
+        return 0
+
+    def control_merge(self, control, control_prev, output_dtype):
+        out = {'input':[], 'middle':[], 'output': []}
+
+        for key in control:
+            control_output = control[key]
+            applied_to = set()
+            for i in range(len(control_output)):
+                x = control_output[i]
+                if x is not None:
+                    if self.global_average_pooling:
+                        x = torch.mean(x, dim=(2, 3), keepdim=True).repeat(1, 1, x.shape[2], x.shape[3])
+
+                    if x not in applied_to: #memory saving strategy, allow shared tensors and only apply strength to shared tensors once
+                        applied_to.add(x)
+                        x *= self.strength
+
+                    if x.dtype != output_dtype:
+                        x = x.to(output_dtype)
+
+                out[key].append(x)
+
+        if control_prev is not None:
+            for x in ['input', 'middle', 'output']:
+                o = out[x]
+                for i in range(len(control_prev[x])):
+                    prev_val = control_prev[x][i]
+                    if i >= len(o):
+                        o.append(prev_val)
+                    elif prev_val is not None:
+                        if o[i] is None:
+                            o[i] = prev_val
+                        else:
+                            if o[i].shape[0] < prev_val.shape[0]:
+                                o[i] = prev_val + o[i]
+                            else:
+                                o[i] = prev_val + o[i] #TODO: change back to inplace add if shared tensors stop being an issue
+        return out
+
+    def set_extra_arg(self, argument, value=None):
+        self.extra_args[argument] = value
+
+
+class ControlNet(ControlBase):
+    def __init__(self, control_model=None, global_average_pooling=False, compression_ratio=8, latent_format=None, device=None, load_device=None, manual_cast_dtype=None):
+        super().__init__(device)
+        self.control_model = control_model
+        self.load_device = load_device
+        if control_model is not None:
+            self.control_model_wrapped = comfy.model_patcher.ModelPatcher(self.control_model, load_device=load_device, offload_device=comfy.model_management.unet_offload_device())
+
+        self.compression_ratio = compression_ratio
+        self.global_average_pooling = global_average_pooling
+        self.model_sampling_current = None
+        self.manual_cast_dtype = manual_cast_dtype
+        self.latent_format = latent_format
+
+    def get_control(self, x_noisy, t, cond, batched_number):
+        control_prev = None
+        if self.previous_controlnet is not None:
+            control_prev = self.previous_controlnet.get_control(x_noisy, t, cond, batched_number)
+
+        if self.timestep_range is not None:
+            if t[0] > self.timestep_range[0] or t[0] < self.timestep_range[1]:
+                if control_prev is not None:
+                    return control_prev
+                else:
+                    return None
+
+        dtype = self.control_model.dtype
+        if self.manual_cast_dtype is not None:
+            dtype = self.manual_cast_dtype
+
+        output_dtype = x_noisy.dtype
+        if self.cond_hint is None or x_noisy.shape[2] * self.compression_ratio != self.cond_hint.shape[2] or x_noisy.shape[3] * self.compression_ratio != self.cond_hint.shape[3]:
+            if self.cond_hint is not None:
+                del self.cond_hint
+            self.cond_hint = None
+            compression_ratio = self.compression_ratio
+            if self.vae is not None:
+                compression_ratio *= self.vae.downscale_ratio
+            self.cond_hint = comfy.utils.common_upscale(self.cond_hint_original, x_noisy.shape[3] * compression_ratio, x_noisy.shape[2] * compression_ratio, self.upscale_algorithm, "center")
+            if self.vae is not None:
+                loaded_models = comfy.model_management.loaded_models(only_currently_used=True)
+                self.cond_hint = self.vae.encode(self.cond_hint.movedim(1, -1))
+                comfy.model_management.load_models_gpu(loaded_models)
+            if self.latent_format is not None:
+                self.cond_hint = self.latent_format.process_in(self.cond_hint)
+            self.cond_hint = self.cond_hint.to(device=self.device, dtype=dtype)
+        if x_noisy.shape[0] != self.cond_hint.shape[0]:
+            self.cond_hint = broadcast_image_to(self.cond_hint, x_noisy.shape[0], batched_number)
+
+        context = cond.get('crossattn_controlnet', cond['c_crossattn'])
+        extra = self.extra_args.copy()
+        for c in ["y", "guidance"]: #TODO
+            temp = cond.get(c, None)
+            if temp is not None:
+                extra[c] = temp.to(dtype)
+
+        timestep = self.model_sampling_current.timestep(t)
+        x_noisy = self.model_sampling_current.calculate_input(t, x_noisy)
+
+        control = self.control_model(x=x_noisy.to(dtype), hint=self.cond_hint, timesteps=timestep.to(dtype), context=context.to(dtype), **extra)
+        return self.control_merge(control, control_prev, output_dtype)
+
+    def copy(self):
+        c = ControlNet(None, global_average_pooling=self.global_average_pooling, load_device=self.load_device, manual_cast_dtype=self.manual_cast_dtype)
+        c.control_model = self.control_model
+        c.control_model_wrapped = self.control_model_wrapped
+        self.copy_to(c)
+        return c
+
+    def get_models(self):
+        out = super().get_models()
+        out.append(self.control_model_wrapped)
+        return out
+
+    def pre_run(self, model, percent_to_timestep_function):
+        super().pre_run(model, percent_to_timestep_function)
+        self.model_sampling_current = model.model_sampling
+
+    def cleanup(self):
+        self.model_sampling_current = None
+        super().cleanup()
+
+class ControlLoraOps:
+    class Linear(torch.nn.Module, comfy.ops.CastWeightBiasOp):
+        def __init__(self, in_features: int, out_features: int, bias: bool = True,
+                    device=None, dtype=None) -> None:
+            factory_kwargs = {'device': device, 'dtype': dtype}
+            super().__init__()
+            self.in_features = in_features
+            self.out_features = out_features
+            self.weight = None
+            self.up = None
+            self.down = None
+            self.bias = None
+
+        def forward(self, input):
+            weight, bias = comfy.ops.cast_bias_weight(self, input)
+            if self.up is not None:
+                return torch.nn.functional.linear(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias)
+            else:
+                return torch.nn.functional.linear(input, weight, bias)
+
+    class Conv2d(torch.nn.Module, comfy.ops.CastWeightBiasOp):
+        def __init__(
+            self,
+            in_channels,
+            out_channels,
+            kernel_size,
+            stride=1,
+            padding=0,
+            dilation=1,
+            groups=1,
+            bias=True,
+            padding_mode='zeros',
+            device=None,
+            dtype=None
+        ):
+            super().__init__()
+            self.in_channels = in_channels
+            self.out_channels = out_channels
+            self.kernel_size = kernel_size
+            self.stride = stride
+            self.padding = padding
+            self.dilation = dilation
+            self.transposed = False
+            self.output_padding = 0
+            self.groups = groups
+            self.padding_mode = padding_mode
+
+            self.weight = None
+            self.bias = None
+            self.up = None
+            self.down = None
+
+
+        def forward(self, input):
+            weight, bias = comfy.ops.cast_bias_weight(self, input)
+            if self.up is not None:
+                return torch.nn.functional.conv2d(input, weight + (torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1))).reshape(self.weight.shape).type(input.dtype), bias, self.stride, self.padding, self.dilation, self.groups)
+            else:
+                return torch.nn.functional.conv2d(input, weight, bias, self.stride, self.padding, self.dilation, self.groups)
+
+
+class ControlLora(ControlNet):
+    def __init__(self, control_weights, global_average_pooling=False, device=None):
+        ControlBase.__init__(self, device)
+        self.control_weights = control_weights
+        self.global_average_pooling = global_average_pooling
+
+    def pre_run(self, model, percent_to_timestep_function):
+        super().pre_run(model, percent_to_timestep_function)
+        controlnet_config = model.model_config.unet_config.copy()
+        controlnet_config.pop("out_channels")
+        controlnet_config["hint_channels"] = self.control_weights["input_hint_block.0.weight"].shape[1]
+        self.manual_cast_dtype = model.manual_cast_dtype
+        dtype = model.get_dtype()
+        if self.manual_cast_dtype is None:
+            class control_lora_ops(ControlLoraOps, comfy.ops.disable_weight_init):
+                pass
+        else:
+            class control_lora_ops(ControlLoraOps, comfy.ops.manual_cast):
+                pass
+            dtype = self.manual_cast_dtype
+
+        controlnet_config["operations"] = control_lora_ops
+        controlnet_config["dtype"] = dtype
+        self.control_model = comfy.cldm.cldm.ControlNet(**controlnet_config)
+        self.control_model.to(comfy.model_management.get_torch_device())
+        diffusion_model = model.diffusion_model
+        sd = diffusion_model.state_dict()
+        cm = self.control_model.state_dict()
+
+        for k in sd:
+            weight = sd[k]
+            try:
+                comfy.utils.set_attr_param(self.control_model, k, weight)
+            except:
+                pass
+
+        for k in self.control_weights:
+            if k not in {"lora_controlnet"}:
+                comfy.utils.set_attr_param(self.control_model, k, self.control_weights[k].to(dtype).to(comfy.model_management.get_torch_device()))
+
+    def copy(self):
+        c = ControlLora(self.control_weights, global_average_pooling=self.global_average_pooling)
+        self.copy_to(c)
+        return c
+
+    def cleanup(self):
+        del self.control_model
+        self.control_model = None
+        super().cleanup()
+
+    def get_models(self):
+        out = ControlBase.get_models(self)
+        return out
+
+    def inference_memory_requirements(self, dtype):
+        return comfy.utils.calculate_parameters(self.control_weights) * comfy.model_management.dtype_size(dtype) + ControlBase.inference_memory_requirements(self, dtype)
+
+def controlnet_config(sd):
+    model_config = comfy.model_detection.model_config_from_unet(sd, "", True)
+
+    supported_inference_dtypes = model_config.supported_inference_dtypes
+
+    controlnet_config = model_config.unet_config
+    unet_dtype = comfy.model_management.unet_dtype(supported_dtypes=supported_inference_dtypes)
+    load_device = comfy.model_management.get_torch_device()
+    manual_cast_dtype = comfy.model_management.unet_manual_cast(unet_dtype, load_device)
+    if manual_cast_dtype is not None:
+        operations = comfy.ops.manual_cast
+    else:
+        operations = comfy.ops.disable_weight_init
+
+    return model_config, operations, load_device, unet_dtype, manual_cast_dtype
+
+def controlnet_load_state_dict(control_model, sd):
+    missing, unexpected = control_model.load_state_dict(sd, strict=False)
+
+    if len(missing) > 0:
+        logging.warning("missing controlnet keys: {}".format(missing))
+
+    if len(unexpected) > 0:
+        logging.debug("unexpected controlnet keys: {}".format(unexpected))
+    return control_model
+
+def load_controlnet_mmdit(sd):
+    new_sd = comfy.model_detection.convert_diffusers_mmdit(sd, "")
+    model_config, operations, load_device, unet_dtype, manual_cast_dtype = controlnet_config(new_sd)
+    num_blocks = comfy.model_detection.count_blocks(new_sd, 'joint_blocks.{}.')
+    for k in sd:
+        new_sd[k] = sd[k]
+
+    control_model = comfy.cldm.mmdit.ControlNet(num_blocks=num_blocks, operations=operations, device=load_device, dtype=unet_dtype, **model_config.unet_config)
+    control_model = controlnet_load_state_dict(control_model, new_sd)
+
+    latent_format = comfy.latent_formats.SD3()
+    latent_format.shift_factor = 0 #SD3 controlnet weirdness
+    control = ControlNet(control_model, compression_ratio=1, latent_format=latent_format, load_device=load_device, manual_cast_dtype=manual_cast_dtype)
+    return control
+
+
+def load_controlnet(ckpt_path, model=None):
+    controlnet_data = comfy.utils.load_torch_file(ckpt_path, safe_load=True)
+    if "lora_controlnet" in controlnet_data:
+        return ControlLora(controlnet_data)
+
+    controlnet_config = None
+    supported_inference_dtypes = None
+
+    if "controlnet_cond_embedding.conv_in.weight" in controlnet_data: #diffusers format
+        controlnet_config = comfy.model_detection.unet_config_from_diffusers_unet(controlnet_data)
+        diffusers_keys = comfy.utils.unet_to_diffusers(controlnet_config)
+        diffusers_keys["controlnet_mid_block.weight"] = "middle_block_out.0.weight"
+        diffusers_keys["controlnet_mid_block.bias"] = "middle_block_out.0.bias"
+
+        count = 0
+        loop = True
+        while loop:
+            suffix = [".weight", ".bias"]
+            for s in suffix:
+                k_in = "controlnet_down_blocks.{}{}".format(count, s)
+                k_out = "zero_convs.{}.0{}".format(count, s)
+                if k_in not in controlnet_data:
+                    loop = False
+                    break
+                diffusers_keys[k_in] = k_out
+            count += 1
+
+        count = 0
+        loop = True
+        while loop:
+            suffix = [".weight", ".bias"]
+            for s in suffix:
+                if count == 0:
+                    k_in = "controlnet_cond_embedding.conv_in{}".format(s)
+                else:
+                    k_in = "controlnet_cond_embedding.blocks.{}{}".format(count - 1, s)
+                k_out = "input_hint_block.{}{}".format(count * 2, s)
+                if k_in not in controlnet_data:
+                    k_in = "controlnet_cond_embedding.conv_out{}".format(s)
+                    loop = False
+                diffusers_keys[k_in] = k_out
+            count += 1
+
+        new_sd = {}
+        for k in diffusers_keys:
+            if k in controlnet_data:
+                new_sd[diffusers_keys[k]] = controlnet_data.pop(k)
+
+        if "control_add_embedding.linear_1.bias" in controlnet_data: #Union Controlnet
+            controlnet_config["union_controlnet_num_control_type"] = controlnet_data["task_embedding"].shape[0]
+            for k in list(controlnet_data.keys()):
+                new_k = k.replace('.attn.in_proj_', '.attn.in_proj.')
+                new_sd[new_k] = controlnet_data.pop(k)
+
+        leftover_keys = controlnet_data.keys()
+        if len(leftover_keys) > 0:
+            logging.warning("leftover keys: {}".format(leftover_keys))
+        controlnet_data = new_sd
+    elif "controlnet_blocks.0.weight" in controlnet_data: #SD3 diffusers format
+        return load_controlnet_mmdit(controlnet_data)
+
+    pth_key = 'control_model.zero_convs.0.0.weight'
+    pth = False
+    key = 'zero_convs.0.0.weight'
+    if pth_key in controlnet_data:
+        pth = True
+        key = pth_key
+        prefix = "control_model."
+    elif key in controlnet_data:
+        prefix = ""
+    else:
+        net = load_t2i_adapter(controlnet_data)
+        if net is None:
+            logging.error("error checkpoint does not contain controlnet or t2i adapter data {}".format(ckpt_path))
+        return net
+
+    if controlnet_config is None:
+        model_config = comfy.model_detection.model_config_from_unet(controlnet_data, prefix, True)
+        supported_inference_dtypes = model_config.supported_inference_dtypes
+        controlnet_config = model_config.unet_config
+
+    load_device = comfy.model_management.get_torch_device()
+    if supported_inference_dtypes is None:
+        unet_dtype = comfy.model_management.unet_dtype()
+    else:
+        unet_dtype = comfy.model_management.unet_dtype(supported_dtypes=supported_inference_dtypes)
+
+    manual_cast_dtype = comfy.model_management.unet_manual_cast(unet_dtype, load_device)
+    if manual_cast_dtype is not None:
+        controlnet_config["operations"] = comfy.ops.manual_cast
+    controlnet_config["dtype"] = unet_dtype
+    controlnet_config.pop("out_channels")
+    controlnet_config["hint_channels"] = controlnet_data["{}input_hint_block.0.weight".format(prefix)].shape[1]
+    control_model = comfy.cldm.cldm.ControlNet(**controlnet_config)
+
+    if pth:
+        if 'difference' in controlnet_data:
+            if model is not None:
+                comfy.model_management.load_models_gpu([model])
+                model_sd = model.model_state_dict()
+                for x in controlnet_data:
+                    c_m = "control_model."
+                    if x.startswith(c_m):
+                        sd_key = "diffusion_model.{}".format(x[len(c_m):])
+                        if sd_key in model_sd:
+                            cd = controlnet_data[x]
+                            cd += model_sd[sd_key].type(cd.dtype).to(cd.device)
+            else:
+                logging.warning("WARNING: Loaded a diff controlnet without a model. It will very likely not work.")
+
+        class WeightsLoader(torch.nn.Module):
+            pass
+        w = WeightsLoader()
+        w.control_model = control_model
+        missing, unexpected = w.load_state_dict(controlnet_data, strict=False)
+    else:
+        missing, unexpected = control_model.load_state_dict(controlnet_data, strict=False)
+
+    if len(missing) > 0:
+        logging.warning("missing controlnet keys: {}".format(missing))
+
+    if len(unexpected) > 0:
+        logging.debug("unexpected controlnet keys: {}".format(unexpected))
+
+    global_average_pooling = False
+    filename = os.path.splitext(ckpt_path)[0]
+    if filename.endswith("_shuffle") or filename.endswith("_shuffle_fp16"): #TODO: smarter way of enabling global_average_pooling
+        global_average_pooling = True
+
+    control = ControlNet(control_model, global_average_pooling=global_average_pooling, load_device=load_device, manual_cast_dtype=manual_cast_dtype)
+    return control
+
+class T2IAdapter(ControlBase):
+    def __init__(self, t2i_model, channels_in, compression_ratio, upscale_algorithm, device=None):
+        super().__init__(device)
+        self.t2i_model = t2i_model
+        self.channels_in = channels_in
+        self.control_input = None
+        self.compression_ratio = compression_ratio
+        self.upscale_algorithm = upscale_algorithm
+
+    def scale_image_to(self, width, height):
+        unshuffle_amount = self.t2i_model.unshuffle_amount
+        width = math.ceil(width / unshuffle_amount) * unshuffle_amount
+        height = math.ceil(height / unshuffle_amount) * unshuffle_amount
+        return width, height
+
+    def get_control(self, x_noisy, t, cond, batched_number):
+        control_prev = None
+        if self.previous_controlnet is not None:
+            control_prev = self.previous_controlnet.get_control(x_noisy, t, cond, batched_number)
+
+        if self.timestep_range is not None:
+            if t[0] > self.timestep_range[0] or t[0] < self.timestep_range[1]:
+                if control_prev is not None:
+                    return control_prev
+                else:
+                    return None
+
+        if self.cond_hint is None or x_noisy.shape[2] * self.compression_ratio != self.cond_hint.shape[2] or x_noisy.shape[3] * self.compression_ratio != self.cond_hint.shape[3]:
+            if self.cond_hint is not None:
+                del self.cond_hint
+            self.control_input = None
+            self.cond_hint = None
+            width, height = self.scale_image_to(x_noisy.shape[3] * self.compression_ratio, x_noisy.shape[2] * self.compression_ratio)
+            self.cond_hint = comfy.utils.common_upscale(self.cond_hint_original, width, height, self.upscale_algorithm, "center").float().to(self.device)
+            if self.channels_in == 1 and self.cond_hint.shape[1] > 1:
+                self.cond_hint = torch.mean(self.cond_hint, 1, keepdim=True)
+        if x_noisy.shape[0] != self.cond_hint.shape[0]:
+            self.cond_hint = broadcast_image_to(self.cond_hint, x_noisy.shape[0], batched_number)
+        if self.control_input is None:
+            self.t2i_model.to(x_noisy.dtype)
+            self.t2i_model.to(self.device)
+            self.control_input = self.t2i_model(self.cond_hint.to(x_noisy.dtype))
+            self.t2i_model.cpu()
+
+        control_input = {}
+        for k in self.control_input:
+            control_input[k] = list(map(lambda a: None if a is None else a.clone(), self.control_input[k]))
+
+        return self.control_merge(control_input, control_prev, x_noisy.dtype)
+
+    def copy(self):
+        c = T2IAdapter(self.t2i_model, self.channels_in, self.compression_ratio, self.upscale_algorithm)
+        self.copy_to(c)
+        return c
+
+def load_t2i_adapter(t2i_data):
+    compression_ratio = 8
+    upscale_algorithm = 'nearest-exact'
+
+    if 'adapter' in t2i_data:
+        t2i_data = t2i_data['adapter']
+    if 'adapter.body.0.resnets.0.block1.weight' in t2i_data: #diffusers format
+        prefix_replace = {}
+        for i in range(4):
+            for j in range(2):
+                prefix_replace["adapter.body.{}.resnets.{}.".format(i, j)] = "body.{}.".format(i * 2 + j)
+            prefix_replace["adapter.body.{}.".format(i, j)] = "body.{}.".format(i * 2)
+        prefix_replace["adapter."] = ""
+        t2i_data = comfy.utils.state_dict_prefix_replace(t2i_data, prefix_replace)
+    keys = t2i_data.keys()
+
+    if "body.0.in_conv.weight" in keys:
+        cin = t2i_data['body.0.in_conv.weight'].shape[1]
+        model_ad = comfy.t2i_adapter.adapter.Adapter_light(cin=cin, channels=[320, 640, 1280, 1280], nums_rb=4)
+    elif 'conv_in.weight' in keys:
+        cin = t2i_data['conv_in.weight'].shape[1]
+        channel = t2i_data['conv_in.weight'].shape[0]
+        ksize = t2i_data['body.0.block2.weight'].shape[2]
+        use_conv = False
+        down_opts = list(filter(lambda a: a.endswith("down_opt.op.weight"), keys))
+        if len(down_opts) > 0:
+            use_conv = True
+        xl = False
+        if cin == 256 or cin == 768:
+            xl = True
+        model_ad = comfy.t2i_adapter.adapter.Adapter(cin=cin, channels=[channel, channel*2, channel*4, channel*4][:4], nums_rb=2, ksize=ksize, sk=True, use_conv=use_conv, xl=xl)
+    elif "backbone.0.0.weight" in keys:
+        model_ad = comfy.ldm.cascade.controlnet.ControlNet(c_in=t2i_data['backbone.0.0.weight'].shape[1], proj_blocks=[0, 4, 8, 12, 51, 55, 59, 63])
+        compression_ratio = 32
+        upscale_algorithm = 'bilinear'
+    elif "backbone.10.blocks.0.weight" in keys:
+        model_ad = comfy.ldm.cascade.controlnet.ControlNet(c_in=t2i_data['backbone.0.weight'].shape[1], bottleneck_mode="large", proj_blocks=[0, 4, 8, 12, 51, 55, 59, 63])
+        compression_ratio = 1
+        upscale_algorithm = 'nearest-exact'
+    else:
+        return None
+
+    missing, unexpected = model_ad.load_state_dict(t2i_data)
+    if len(missing) > 0:
+        logging.warning("t2i missing {}".format(missing))
+
+    if len(unexpected) > 0:
+        logging.debug("t2i unexpected {}".format(unexpected))
+
+    return T2IAdapter(model_ad, model_ad.input_channels, compression_ratio, upscale_algorithm)

ComfyUI/comfy/diffusers_convert.py

@@ -0,0 +1,281 @@
+import re
+import torch
+import logging
+
+# conversion code from https://github.com/huggingface/diffusers/blob/main/scripts/convert_diffusers_to_original_stable_diffusion.py
+
+# =================#
+# UNet Conversion #
+# =================#
+
+unet_conversion_map = [
+    # (stable-diffusion, HF Diffusers)
+    ("time_embed.0.weight", "time_embedding.linear_1.weight"),
+    ("time_embed.0.bias", "time_embedding.linear_1.bias"),
+    ("time_embed.2.weight", "time_embedding.linear_2.weight"),
+    ("time_embed.2.bias", "time_embedding.linear_2.bias"),
+    ("input_blocks.0.0.weight", "conv_in.weight"),
+    ("input_blocks.0.0.bias", "conv_in.bias"),
+    ("out.0.weight", "conv_norm_out.weight"),
+    ("out.0.bias", "conv_norm_out.bias"),
+    ("out.2.weight", "conv_out.weight"),
+    ("out.2.bias", "conv_out.bias"),
+]
+
+unet_conversion_map_resnet = [
+    # (stable-diffusion, HF Diffusers)
+    ("in_layers.0", "norm1"),
+    ("in_layers.2", "conv1"),
+    ("out_layers.0", "norm2"),
+    ("out_layers.3", "conv2"),
+    ("emb_layers.1", "time_emb_proj"),
+    ("skip_connection", "conv_shortcut"),
+]
+
+unet_conversion_map_layer = []
+# hardcoded number of downblocks and resnets/attentions...
+# would need smarter logic for other networks.
+for i in range(4):
+    # loop over downblocks/upblocks
+
+    for j in range(2):
+        # loop over resnets/attentions for downblocks
+        hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}."
+        sd_down_res_prefix = f"input_blocks.{3 * i + j + 1}.0."
+        unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix))
+
+        if i < 3:
+            # no attention layers in down_blocks.3
+            hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}."
+            sd_down_atn_prefix = f"input_blocks.{3 * i + j + 1}.1."
+            unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix))
+
+    for j in range(3):
+        # loop over resnets/attentions for upblocks
+        hf_up_res_prefix = f"up_blocks.{i}.resnets.{j}."
+        sd_up_res_prefix = f"output_blocks.{3 * i + j}.0."
+        unet_conversion_map_layer.append((sd_up_res_prefix, hf_up_res_prefix))
+
+        if i > 0:
+            # no attention layers in up_blocks.0
+            hf_up_atn_prefix = f"up_blocks.{i}.attentions.{j}."
+            sd_up_atn_prefix = f"output_blocks.{3 * i + j}.1."
+            unet_conversion_map_layer.append((sd_up_atn_prefix, hf_up_atn_prefix))
+
+    if i < 3:
+        # no downsample in down_blocks.3
+        hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv."
+        sd_downsample_prefix = f"input_blocks.{3 * (i + 1)}.0.op."
+        unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix))
+
+        # no upsample in up_blocks.3
+        hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
+        sd_upsample_prefix = f"output_blocks.{3 * i + 2}.{1 if i == 0 else 2}."
+        unet_conversion_map_layer.append((sd_upsample_prefix, hf_upsample_prefix))
+
+hf_mid_atn_prefix = "mid_block.attentions.0."
+sd_mid_atn_prefix = "middle_block.1."
+unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix))
+
+for j in range(2):
+    hf_mid_res_prefix = f"mid_block.resnets.{j}."
+    sd_mid_res_prefix = f"middle_block.{2 * j}."
+    unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix))
+
+
+def convert_unet_state_dict(unet_state_dict):
+    # buyer beware: this is a *brittle* function,
+    # and correct output requires that all of these pieces interact in
+    # the exact order in which I have arranged them.
+    mapping = {k: k for k in unet_state_dict.keys()}
+    for sd_name, hf_name in unet_conversion_map:
+        mapping[hf_name] = sd_name
+    for k, v in mapping.items():
+        if "resnets" in k:
+            for sd_part, hf_part in unet_conversion_map_resnet:
+                v = v.replace(hf_part, sd_part)
+            mapping[k] = v
+    for k, v in mapping.items():
+        for sd_part, hf_part in unet_conversion_map_layer:
+            v = v.replace(hf_part, sd_part)
+        mapping[k] = v
+    new_state_dict = {v: unet_state_dict[k] for k, v in mapping.items()}
+    return new_state_dict
+
+
+# ================#
+# VAE Conversion #
+# ================#
+
+vae_conversion_map = [
+    # (stable-diffusion, HF Diffusers)
+    ("nin_shortcut", "conv_shortcut"),
+    ("norm_out", "conv_norm_out"),
+    ("mid.attn_1.", "mid_block.attentions.0."),
+]
+
+for i in range(4):
+    # down_blocks have two resnets
+    for j in range(2):
+        hf_down_prefix = f"encoder.down_blocks.{i}.resnets.{j}."
+        sd_down_prefix = f"encoder.down.{i}.block.{j}."
+        vae_conversion_map.append((sd_down_prefix, hf_down_prefix))
+
+    if i < 3:
+        hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0."
+        sd_downsample_prefix = f"down.{i}.downsample."
+        vae_conversion_map.append((sd_downsample_prefix, hf_downsample_prefix))
+
+        hf_upsample_prefix = f"up_blocks.{i}.upsamplers.0."
+        sd_upsample_prefix = f"up.{3 - i}.upsample."
+        vae_conversion_map.append((sd_upsample_prefix, hf_upsample_prefix))
+
+    # up_blocks have three resnets
+    # also, up blocks in hf are numbered in reverse from sd
+    for j in range(3):
+        hf_up_prefix = f"decoder.up_blocks.{i}.resnets.{j}."
+        sd_up_prefix = f"decoder.up.{3 - i}.block.{j}."
+        vae_conversion_map.append((sd_up_prefix, hf_up_prefix))
+
+# this part accounts for mid blocks in both the encoder and the decoder
+for i in range(2):
+    hf_mid_res_prefix = f"mid_block.resnets.{i}."
+    sd_mid_res_prefix = f"mid.block_{i + 1}."
+    vae_conversion_map.append((sd_mid_res_prefix, hf_mid_res_prefix))
+
+vae_conversion_map_attn = [
+    # (stable-diffusion, HF Diffusers)
+    ("norm.", "group_norm."),
+    ("q.", "query."),
+    ("k.", "key."),
+    ("v.", "value."),
+    ("q.", "to_q."),
+    ("k.", "to_k."),
+    ("v.", "to_v."),
+    ("proj_out.", "to_out.0."),
+    ("proj_out.", "proj_attn."),
+]
+
+
+def reshape_weight_for_sd(w):
+    # convert HF linear weights to SD conv2d weights
+    return w.reshape(*w.shape, 1, 1)
+
+
+def convert_vae_state_dict(vae_state_dict):
+    mapping = {k: k for k in vae_state_dict.keys()}
+    for k, v in mapping.items():
+        for sd_part, hf_part in vae_conversion_map:
+            v = v.replace(hf_part, sd_part)
+        mapping[k] = v
+    for k, v in mapping.items():
+        if "attentions" in k:
+            for sd_part, hf_part in vae_conversion_map_attn:
+                v = v.replace(hf_part, sd_part)
+            mapping[k] = v
+    new_state_dict = {v: vae_state_dict[k] for k, v in mapping.items()}
+    weights_to_convert = ["q", "k", "v", "proj_out"]
+    for k, v in new_state_dict.items():
+        for weight_name in weights_to_convert:
+            if f"mid.attn_1.{weight_name}.weight" in k:
+                logging.debug(f"Reshaping {k} for SD format")
+                new_state_dict[k] = reshape_weight_for_sd(v)
+    return new_state_dict
+
+
+# =========================#
+# Text Encoder Conversion #
+# =========================#
+
+
+textenc_conversion_lst = [
+    # (stable-diffusion, HF Diffusers)
+    ("resblocks.", "text_model.encoder.layers."),
+    ("ln_1", "layer_norm1"),
+    ("ln_2", "layer_norm2"),
+    (".c_fc.", ".fc1."),
+    (".c_proj.", ".fc2."),
+    (".attn", ".self_attn"),
+    ("ln_final.", "transformer.text_model.final_layer_norm."),
+    ("token_embedding.weight", "transformer.text_model.embeddings.token_embedding.weight"),
+    ("positional_embedding", "transformer.text_model.embeddings.position_embedding.weight"),
+]
+protected = {re.escape(x[1]): x[0] for x in textenc_conversion_lst}
+textenc_pattern = re.compile("|".join(protected.keys()))
+
+# Ordering is from https://github.com/pytorch/pytorch/blob/master/test/cpp/api/modules.cpp
+code2idx = {"q": 0, "k": 1, "v": 2}
+
+# This function exists because at the time of writing torch.cat can't do fp8 with cuda
+def cat_tensors(tensors):
+    x = 0
+    for t in tensors:
+        x += t.shape[0]
+
+    shape = [x] + list(tensors[0].shape)[1:]
+    out = torch.empty(shape, device=tensors[0].device, dtype=tensors[0].dtype)
+
+    x = 0
+    for t in tensors:
+        out[x:x + t.shape[0]] = t
+        x += t.shape[0]
+
+    return out
+
+def convert_text_enc_state_dict_v20(text_enc_dict, prefix=""):
+    new_state_dict = {}
+    capture_qkv_weight = {}
+    capture_qkv_bias = {}
+    for k, v in text_enc_dict.items():
+        if not k.startswith(prefix):
+            continue
+        if (
+                k.endswith(".self_attn.q_proj.weight")
+                or k.endswith(".self_attn.k_proj.weight")
+                or k.endswith(".self_attn.v_proj.weight")
+        ):
+            k_pre = k[: -len(".q_proj.weight")]
+            k_code = k[-len("q_proj.weight")]
+            if k_pre not in capture_qkv_weight:
+                capture_qkv_weight[k_pre] = [None, None, None]
+            capture_qkv_weight[k_pre][code2idx[k_code]] = v
+            continue
+
+        if (
+                k.endswith(".self_attn.q_proj.bias")
+                or k.endswith(".self_attn.k_proj.bias")
+                or k.endswith(".self_attn.v_proj.bias")
+        ):
+            k_pre = k[: -len(".q_proj.bias")]
+            k_code = k[-len("q_proj.bias")]
+            if k_pre not in capture_qkv_bias:
+                capture_qkv_bias[k_pre] = [None, None, None]
+            capture_qkv_bias[k_pre][code2idx[k_code]] = v
+            continue
+
+        text_proj = "transformer.text_projection.weight"
+        if k.endswith(text_proj):
+            new_state_dict[k.replace(text_proj, "text_projection")] = v.transpose(0, 1).contiguous()
+        else:
+            relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k)
+            new_state_dict[relabelled_key] = v
+
+    for k_pre, tensors in capture_qkv_weight.items():
+        if None in tensors:
+            raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")
+        relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)
+        new_state_dict[relabelled_key + ".in_proj_weight"] = cat_tensors(tensors)
+
+    for k_pre, tensors in capture_qkv_bias.items():
+        if None in tensors:
+            raise Exception("CORRUPTED MODEL: one of the q-k-v values for the text encoder was missing")
+        relabelled_key = textenc_pattern.sub(lambda m: protected[re.escape(m.group(0))], k_pre)
+        new_state_dict[relabelled_key + ".in_proj_bias"] = cat_tensors(tensors)
+
+    return new_state_dict
+
+
+def convert_text_enc_state_dict(text_enc_dict):
+    return text_enc_dict
+
+

ComfyUI/comfy/diffusers_load.py

@@ -0,0 +1,36 @@
+import os
+
+import comfy.sd
+
+def first_file(path, filenames):
+    for f in filenames:
+        p = os.path.join(path, f)
+        if os.path.exists(p):
+            return p
+    return None
+
+def load_diffusers(model_path, output_vae=True, output_clip=True, embedding_directory=None):
+    diffusion_model_names = ["diffusion_pytorch_model.fp16.safetensors", "diffusion_pytorch_model.safetensors", "diffusion_pytorch_model.fp16.bin", "diffusion_pytorch_model.bin"]
+    unet_path = first_file(os.path.join(model_path, "unet"), diffusion_model_names)
+    vae_path = first_file(os.path.join(model_path, "vae"), diffusion_model_names)
+
+    text_encoder_model_names = ["model.fp16.safetensors", "model.safetensors", "pytorch_model.fp16.bin", "pytorch_model.bin"]
+    text_encoder1_path = first_file(os.path.join(model_path, "text_encoder"), text_encoder_model_names)
+    text_encoder2_path = first_file(os.path.join(model_path, "text_encoder_2"), text_encoder_model_names)
+
+    text_encoder_paths = [text_encoder1_path]
+    if text_encoder2_path is not None:
+        text_encoder_paths.append(text_encoder2_path)
+
+    unet = comfy.sd.load_unet(unet_path)
+
+    clip = None
+    if output_clip:
+        clip = comfy.sd.load_clip(text_encoder_paths, embedding_directory=embedding_directory)
+
+    vae = None
+    if output_vae:
+        sd = comfy.utils.load_torch_file(vae_path)
+        vae = comfy.sd.VAE(sd=sd)
+
+    return (unet, clip, vae)