Merge PR #69 from Kosinkadink/develop - SVD-ControlNet support

SVD-ControlNet support

README.md

@@ -1,5 +1,5 @@
 # ComfyUI-Advanced-ControlNet
-Nodes for scheduling ControlNet strength across timesteps and batched latents, as well as applying custom weights and attention masks. The ControlNet nodes here fully support sliding context sampling, like the one used in the  [ComfyUI-AnimateDiff-Evolved](https://github.com/Kosinkadink/ComfyUI-AnimateDiff-Evolved) nodes. Currently supports ControlNets, T2IAdapters, ControlLoRAs, ControlLLLite, and SparseCtrls.
+Nodes for scheduling ControlNet strength across timesteps and batched latents, as well as applying custom weights and attention masks. The ControlNet nodes here fully support sliding context sampling, like the one used in the  [ComfyUI-AnimateDiff-Evolved](https://github.com/Kosinkadink/ComfyUI-AnimateDiff-Evolved) nodes. Currently supports ControlNets, T2IAdapters, ControlLoRAs, ControlLLLite, SparseCtrls, and SVD-ControlNets.
 
 Custom weights allow replication of the "My prompt is more important" feature of Auto1111's sd-webui ControlNet extension.
 
@@ -12,6 +12,8 @@ ControlNet preprocessors are available through [comfyui_controlnet_aux](https://
 - ControlNet, T2IAdapter, and ControlLoRA support for sliding context windows.
 - ControlLLLite support (requires model_optional to be passed into and out of Apply Advanced ControlNet node)
 - SparseCtrl support
+- SVD-ControlNet support
+  - Stable Video Diffusion ControlNets trained by **CiaraRowles**: [Depth](https://huggingface.co/CiaraRowles/temporal-controlnet-depth-svd-v1/tree/main/controlnet), [Lineart](https://huggingface.co/CiaraRowles/temporal-controlnet-lineart-svd-v1/tree/main/controlnet)  
 
 ## Table of Contents:
 - [Scheduling Explanation](#scheduling-explanation)

adv_control/control.py

@@ -12,6 +12,7 @@ from model_patcher import ModelPatcher
 
 from .control_sparsectrl import SparseControlNet, SparseCtrlMotionWrapper, SparseMethod, SparseSettings, SparseSpreadMethod, PreprocSparseRGBWrapper
 from .control_lllite import LLLiteModule, LLLitePatch
+from .control_svd import svd_unet_config_from_diffusers_unet, SVDControlNet, svd_unet_to_diffusers
 from .utils import (AdvancedControlBase, TimestepKeyframeGroup, LatentKeyframeGroup, ControlWeightType, ControlWeights, WeightTypeException,
                     manual_cast_clean_groupnorm, disable_weight_init_clean_groupnorm, prepare_mask_batch, get_properly_arranged_t2i_weights, load_torch_file_with_dict_factory)
 from .logger import logger
@@ -168,11 +169,70 @@ class ControlLoraAdvanced(ControlLora, AdvancedControlBase):
                                    global_average_pooling=v.global_average_pooling, device=v.device)
 
 
-
 class SVDControlNetAdvanced(ControlNetAdvanced):
-    def __init__(self, control_model, timestep_keyframes: TimestepKeyframeGroup, global_average_pooling=False, device=None, load_device=None, manual_cast_dtype=None):
+    def __init__(self, control_model: SVDControlNet, timestep_keyframes: TimestepKeyframeGroup, global_average_pooling=False, device=None, load_device=None, manual_cast_dtype=None):
         super().__init__(control_model=control_model, timestep_keyframes=timestep_keyframes, global_average_pooling=global_average_pooling, device=device, load_device=load_device, manual_cast_dtype=manual_cast_dtype)
 
+    def set_cond_hint(self, *args, **kwargs):
+        to_return = super().set_cond_hint(*args, **kwargs)
+        # cond hint for SVD-ControlNet needs to be scaled between (-1, 1) instead of (0, 1)
+        self.cond_hint_original = self.cond_hint_original * 2.0 - 1.0
+        return to_return
+
+    def get_control_advanced(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
+        # make cond_hint appropriate dimensions
+        # TODO: change this to not require cond_hint upscaling every step when self.sub_idxs are present
+        if self.sub_idxs is not None or self.cond_hint is None or x_noisy.shape[2] * 8 != self.cond_hint.shape[2] or x_noisy.shape[3] * 8 != self.cond_hint.shape[3]:
+            if self.cond_hint is not None:
+                del self.cond_hint
+            self.cond_hint = None
+            # if self.cond_hint_original length greater or equal to real latent count, subdivide it before scaling
+            if self.sub_idxs is not None and self.cond_hint_original.size(0) >= self.full_latent_length:
+                self.cond_hint = comfy.utils.common_upscale(self.cond_hint_original[self.sub_idxs], x_noisy.shape[3] * 8, x_noisy.shape[2] * 8, 'nearest-exact', "center").to(dtype).to(self.device)
+            else:
+                self.cond_hint = comfy.utils.common_upscale(self.cond_hint_original, x_noisy.shape[3] * 8, x_noisy.shape[2] * 8, 'nearest-exact', "center").to(dtype).to(self.device)
+        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)
+
+        # prepare mask_cond_hint
+        self.prepare_mask_cond_hint(x_noisy=x_noisy, t=t, cond=cond, batched_number=batched_number, dtype=dtype)
+
+        context = cond.get('crossattn_controlnet', cond['c_crossattn'])
+        # uses 'y' in new ComfyUI update
+        y = cond.get('y', None)
+        if y is not None:
+            y = y.to(dtype)
+        timestep = self.model_sampling_current.timestep(t)
+        x_noisy = self.model_sampling_current.calculate_input(t, x_noisy)
+        # concat c_concat if exists (should exist for SVD), doubling channels to 8
+        if cond.get('c_concat', None) is not None:
+            x_noisy = torch.cat([x_noisy] + [cond['c_concat']], dim=1)
+
+        control = self.control_model(x=x_noisy.to(dtype), hint=self.cond_hint, timesteps=timestep.float(), context=context.to(dtype), y=y, cond=cond)
+        return self.control_merge(None, control, control_prev, output_dtype)
+
+    def copy(self):
+        c = SVDControlNetAdvanced(self.control_model, self.timestep_keyframes, global_average_pooling=self.global_average_pooling, load_device=self.load_device, manual_cast_dtype=self.manual_cast_dtype)
+        self.copy_to(c)
+        self.copy_to_advanced(c)
+        return c
+
 
 class SparseCtrlAdvanced(ControlNetAdvanced):
     def __init__(self, control_model, timestep_keyframes: TimestepKeyframeGroup, sparse_settings: SparseSettings=None, global_average_pooling=False, device=None, load_device=None, manual_cast_dtype=None):
@@ -480,7 +540,8 @@ def load_controlnet(ckpt_path, timestep_keyframe: TimestepKeyframeGroup=None, mo
         elif controlnet_type == ControlWeightType.SPARSECTRL:
             control = load_sparsectrl(ckpt_path, controlnet_data=controlnet_data, timestep_keyframe=timestep_keyframe, model=model)
         elif controlnet_type == ControlWeightType.SVD_CONTROLNET:
-            raise Exception(f"SVD-ControlNet is not supported yet!")
+            control = load_svdcontrolnet(ckpt_path, controlnet_data=controlnet_data, timestep_keyframe=timestep_keyframe)
+            #raise Exception(f"SVD-ControlNet is not supported yet!")
             #control = comfy_cn.load_controlnet(ckpt_path, model=model)
     # otherwise, load vanilla ControlNet
     else:
@@ -710,3 +771,125 @@ def load_controllllite(ckpt_path: str, controlnet_data: dict[str, Tensor]=None,
     patch_attn2 = LLLitePatch(modules=modules, patch_type=LLLitePatch.ATTN2)
     control = ControlLLLiteAdvanced(patch_attn1=patch_attn1, patch_attn2=patch_attn2, timestep_keyframes=timestep_keyframe)
     return control
+
+
+def load_svdcontrolnet(ckpt_path: str, controlnet_data: dict[str, Tensor]=None, timestep_keyframe: TimestepKeyframeGroup=None, model=None):
+    if controlnet_data is None:
+        controlnet_data = comfy.utils.load_torch_file(ckpt_path, safe_load=True)
+
+    controlnet_config = None
+    if "controlnet_cond_embedding.conv_in.weight" in controlnet_data: #diffusers format
+        unet_dtype = comfy.model_management.unet_dtype()
+        controlnet_config = svd_unet_config_from_diffusers_unet(controlnet_data, unet_dtype)
+        diffusers_keys = svd_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)
+
+        leftover_keys = controlnet_data.keys()
+        if len(leftover_keys) > 0:
+            spatial_leftover_keys = []
+            temporal_leftover_keys = []
+            other_leftover_keys = []
+            for key in leftover_keys:
+                if "spatial" in key:
+                    spatial_leftover_keys.append(key)
+                elif "temporal" in key:
+                    temporal_leftover_keys.append(key)
+                else:
+                    other_leftover_keys.append(key)
+            logger.warn(f"spatial_leftover_keys ({len(spatial_leftover_keys)}): {spatial_leftover_keys}")
+            logger.warn(f"temporal_leftover_keys ({len(temporal_leftover_keys)}): {temporal_leftover_keys}")
+            logger.warn(f"other_leftover_keys ({len(other_leftover_keys)}): {other_leftover_keys}")
+            #print("leftover keys:", leftover_keys)
+        controlnet_data = new_sd
+
+    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:
+        raise ValueError("The provided model is not a valid SVD-ControlNet model! [ErrorCode: MUSTARD]")
+
+    if controlnet_config is None:
+        unet_dtype = comfy.model_management.unet_dtype()
+        controlnet_config = comfy.model_detection.model_config_from_unet(controlnet_data, prefix, unet_dtype, True).unet_config
+    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:
+        controlnet_config["operations"] = comfy.ops.manual_cast
+    controlnet_config.pop("out_channels")
+    controlnet_config["hint_channels"] = controlnet_data["{}input_hint_block.0.weight".format(prefix)].shape[1]
+    control_model = SVDControlNet(**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:
+                print("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 or len(unexpected) > 0:
+        logger.info(f"SVD-ControlNet: {missing}, {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 = SVDControlNetAdvanced(control_model, timestep_keyframes=timestep_keyframe, global_average_pooling=global_average_pooling, load_device=load_device, manual_cast_dtype=manual_cast_dtype)
+    return control
+

adv_control/control_svd.py

@@ -1,6 +1,517 @@
-from comfy.cldm.cldm import ControlNet as ControlNetCLDM
+import torch
+import torch.nn as nn
+from torch import Tensor
+
+import comfy.model_detection
+from comfy.utils import UNET_MAP_BASIC, UNET_MAP_RESNET, UNET_MAP_ATTENTIONS, TRANSFORMER_BLOCKS
+
+import torch
 
 
-class SVDControlNet(ControlNetCLDM):
-    def __init__(self, *args,**kwargs):
-        super().__init__(*args, **kwargs)
+from comfy.ldm.modules.diffusionmodules.util import (
+    zero_module,
+    timestep_embedding,
+)
+
+from comfy.ldm.modules.attention import SpatialVideoTransformer
+from comfy.ldm.modules.diffusionmodules.openaimodel import UNetModel, TimestepEmbedSequential, VideoResBlock, Downsample 
+from comfy.ldm.util import exists
+import comfy.ops
+
+
+class SVDControlNet(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,
+        use_spatial_context=False,
+        extra_ff_mix_layer=False,
+        merge_strategy="fixed",
+        merge_factor=0.5,
+        video_kernel_size=3,
+        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 = [
+                    VideoResBlock(
+                        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,
+                        video_kernel_size=video_kernel_size,
+                        merge_strategy=merge_strategy, merge_factor=merge_factor,
+                    )
+                ]
+                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(
+                            SpatialVideoTransformer(
+                                ch, num_heads, dim_head, depth=num_transformers, context_dim=context_dim,
+                                disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
+                                checkpoint=use_checkpoint, dtype=self.dtype, device=device, operations=operations,
+                                use_spatial_context=use_spatial_context, ff_in=extra_ff_mix_layer,
+                                merge_strategy=merge_strategy, merge_factor=merge_factor,
+                            )
+                        )
+                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(
+                        VideoResBlock(
+                            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,
+                            video_kernel_size=video_kernel_size,
+                            merge_strategy=merge_strategy, merge_factor=merge_factor,
+                        )
+                        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 = [
+            VideoResBlock(
+                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,
+                video_kernel_size=video_kernel_size,
+                merge_strategy=merge_strategy, merge_factor=merge_factor,
+            )]
+        if transformer_depth_middle >= 0:
+            mid_block += [SpatialVideoTransformer(  # 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,
+                            checkpoint=use_checkpoint, dtype=self.dtype, device=device, operations=operations,
+                            use_spatial_context=use_spatial_context, ff_in=extra_ff_mix_layer,
+                            merge_strategy=merge_strategy, merge_factor=merge_factor,
+                        ),
+            VideoResBlock(
+                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,
+                video_kernel_size=video_kernel_size,
+                merge_strategy=merge_strategy, merge_factor=merge_factor,
+            )]
+        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
+
+    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)
+
+        cond = kwargs["cond"]
+        num_video_frames = cond["num_video_frames"]
+        image_only_indicator = cond["image_only_indicator"]
+        time_context = cond.get("time_context", None)
+        del cond
+
+        guided_hint = self.input_hint_block(hint, emb, context, time_context=time_context, num_video_frames=num_video_frames, image_only_indicator=image_only_indicator)
+
+        outs = []
+
+        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, time_context=time_context, num_video_frames=num_video_frames, image_only_indicator=image_only_indicator)
+                h += guided_hint
+                guided_hint = None
+            else:
+                h = module(h, emb, context, time_context=time_context, num_video_frames=num_video_frames, image_only_indicator=image_only_indicator)
+            outs.append(zero_conv(h, emb, context, time_context=time_context, num_video_frames=num_video_frames, image_only_indicator=image_only_indicator))
+
+        h = self.middle_block(h, emb, context, time_context=time_context, num_video_frames=num_video_frames, image_only_indicator=image_only_indicator)
+        outs.append(self.middle_block_out(h, emb, context, time_context=time_context, num_video_frames=num_video_frames, image_only_indicator=image_only_indicator))
+
+        return outs
+
+
+TEMPORAL_TRANSFORMER_BLOCKS = {
+    "norm_in.weight",
+    "norm_in.bias",
+    "ff_in.net.0.proj.weight",
+    "ff_in.net.0.proj.bias",
+    "ff_in.net.2.weight",
+    "ff_in.net.2.bias",
+}
+TEMPORAL_TRANSFORMER_BLOCKS.update(TRANSFORMER_BLOCKS)
+
+
+TEMPORAL_UNET_MAP_ATTENTIONS = {
+    "time_mixer.mix_factor",
+}
+TEMPORAL_UNET_MAP_ATTENTIONS.update(UNET_MAP_ATTENTIONS)
+
+
+TEMPORAL_TRANSFORMER_MAP = {
+    "time_pos_embed.0.weight": "time_pos_embed.linear_1.weight",
+    "time_pos_embed.0.bias": "time_pos_embed.linear_1.bias",
+    "time_pos_embed.2.weight": "time_pos_embed.linear_2.weight",
+    "time_pos_embed.2.bias": "time_pos_embed.linear_2.bias",
+}
+
+
+TEMPORAL_RESNET = {
+     "time_mixer.mix_factor",
+}
+
+
+def svd_unet_config_from_diffusers_unet(state_dict: dict[str, Tensor], dtype):
+    match = {}
+    transformer_depth = []
+
+    attn_res = 1
+    down_blocks = comfy.model_detection.count_blocks(state_dict, "down_blocks.{}")
+    for i in range(down_blocks):
+        attn_blocks = comfy.model_detection.count_blocks(state_dict, "down_blocks.{}.attentions.".format(i) + '{}')
+        for ab in range(attn_blocks):
+            transformer_count = comfy.model_detection.count_blocks(state_dict, "down_blocks.{}.attentions.{}.transformer_blocks.".format(i, ab) + '{}')
+            transformer_depth.append(transformer_count)
+            if transformer_count > 0:
+                match["context_dim"] = state_dict["down_blocks.{}.attentions.{}.transformer_blocks.0.attn2.to_k.weight".format(i, ab)].shape[1]
+
+        attn_res *= 2
+        if attn_blocks == 0:
+            transformer_depth.append(0)
+            transformer_depth.append(0)
+
+    match["transformer_depth"] = transformer_depth
+
+    match["model_channels"] = state_dict["conv_in.weight"].shape[0]
+    match["in_channels"] = state_dict["conv_in.weight"].shape[1]
+    match["adm_in_channels"] = None
+    if "class_embedding.linear_1.weight" in state_dict:
+        match["adm_in_channels"] = state_dict["class_embedding.linear_1.weight"].shape[1]
+    elif "add_embedding.linear_1.weight" in state_dict:
+        match["adm_in_channels"] = state_dict["add_embedding.linear_1.weight"].shape[1]
+
+    # based on unet_config of SVD
+    SVD = {
+        'use_checkpoint': False,
+        'image_size': 32,
+        'use_spatial_transformer': True,
+        'legacy': False,
+        'num_classes': 'sequential',
+        'adm_in_channels': 768,
+        'dtype': dtype,
+        'in_channels': 8,
+        'out_channels': 4,
+        'model_channels': 320,
+        'num_res_blocks': [2, 2, 2, 2],
+        'transformer_depth': [1, 1, 1, 1, 1, 1, 0, 0],
+        'transformer_depth_output': [1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0],
+        'channel_mult': [1, 2, 4, 4],
+        'transformer_depth_middle': 1,
+        'use_linear_in_transformer': True,
+        'context_dim': 1024,
+        'extra_ff_mix_layer': True,
+        'use_spatial_context': True,
+        'merge_strategy': 'learned_with_images',
+        'merge_factor': 0.0,
+        'video_kernel_size': [3, 1, 1],
+        'use_temporal_attention': True,
+        'use_temporal_resblock': True,
+        'num_heads': -1,
+        'num_head_channels': 64,
+        }
+
+    supported_models = [SVD]
+
+    for unet_config in supported_models:
+        matches = True
+        for k in match:
+            if match[k] != unet_config[k]:
+                matches = False
+                break
+        if matches:
+            return comfy.model_detection.convert_config(unet_config)
+    return None
+
+
+def svd_unet_to_diffusers(unet_config):
+    num_res_blocks = unet_config["num_res_blocks"]
+    channel_mult = unet_config["channel_mult"]
+    transformer_depth = unet_config["transformer_depth"][:]
+    transformer_depth_output = unet_config["transformer_depth_output"][:]
+    num_blocks = len(channel_mult)
+
+    transformers_mid = unet_config.get("transformer_depth_middle", None)
+
+    diffusers_unet_map = {}
+    for x in range(num_blocks):
+        n = 1 + (num_res_blocks[x] + 1) * x
+        for i in range(num_res_blocks[x]):
+            for b in TEMPORAL_RESNET:
+                diffusers_unet_map["down_blocks.{}.resnets.{}.{}".format(x, i, b)] = "input_blocks.{}.0.{}".format(n, b)
+            for b in UNET_MAP_RESNET:
+                diffusers_unet_map["down_blocks.{}.resnets.{}.spatial_res_block.{}".format(x, i, UNET_MAP_RESNET[b])] = "input_blocks.{}.0.{}".format(n, b)
+                diffusers_unet_map["down_blocks.{}.resnets.{}.temporal_res_block.{}".format(x, i, UNET_MAP_RESNET[b])] = "input_blocks.{}.0.time_stack.{}".format(n, b)
+                #diffusers_unet_map["down_blocks.{}.resnets.{}.{}".format(x, i, UNET_MAP_RESNET[b])] = "input_blocks.{}.0.{}".format(n, b)
+            num_transformers = transformer_depth.pop(0)
+            if num_transformers > 0:
+                for b in TEMPORAL_UNET_MAP_ATTENTIONS:
+                    diffusers_unet_map["down_blocks.{}.attentions.{}.{}".format(x, i, b)] = "input_blocks.{}.1.{}".format(n, b)
+                for b in TEMPORAL_TRANSFORMER_MAP:
+                    diffusers_unet_map["down_blocks.{}.attentions.{}.{}".format(x, i, TEMPORAL_TRANSFORMER_MAP[b])] = "input_blocks.{}.1.{}".format(n, b)
+                for t in range(num_transformers):
+                    for b in TRANSFORMER_BLOCKS:
+                        diffusers_unet_map["down_blocks.{}.attentions.{}.transformer_blocks.{}.{}".format(x, i, t, b)] = "input_blocks.{}.1.transformer_blocks.{}.{}".format(n, t, b)
+                    for b in TEMPORAL_TRANSFORMER_BLOCKS:
+                        diffusers_unet_map["down_blocks.{}.attentions.{}.temporal_transformer_blocks.{}.{}".format(x, i, t, b)] = "input_blocks.{}.1.time_stack.{}.{}".format(n, t, b)
+            n += 1
+        for k in ["weight", "bias"]:
+            diffusers_unet_map["down_blocks.{}.downsamplers.0.conv.{}".format(x, k)] = "input_blocks.{}.0.op.{}".format(n, k)
+
+    i = 0
+    for b in TEMPORAL_UNET_MAP_ATTENTIONS:
+        diffusers_unet_map["mid_block.attentions.{}.{}".format(i, b)] = "middle_block.1.{}".format(b)
+    for b in TEMPORAL_TRANSFORMER_MAP:
+        diffusers_unet_map["mid_block.attentions.{}.{}".format(i, TEMPORAL_TRANSFORMER_MAP[b])] = "middle_block.1.{}".format(b)
+    for t in range(transformers_mid):
+        for b in TRANSFORMER_BLOCKS:
+            diffusers_unet_map["mid_block.attentions.{}.transformer_blocks.{}.{}".format(i, t, b)] = "middle_block.1.transformer_blocks.{}.{}".format(t, b)
+        for b in TEMPORAL_TRANSFORMER_BLOCKS:
+            diffusers_unet_map["mid_block.attentions.{}.temporal_transformer_blocks.{}.{}".format(i, t, b)] = "middle_block.1.time_stack.{}.{}".format(t, b)
+
+    for i, n in enumerate([0, 2]):
+        for b in TEMPORAL_RESNET:
+            diffusers_unet_map["mid_block.resnets.{}.{}".format(i, b)] = "middle_block.{}.{}".format(n, b)
+        for b in UNET_MAP_RESNET:
+            diffusers_unet_map["mid_block.resnets.{}.spatial_res_block.{}".format(i, UNET_MAP_RESNET[b])] = "middle_block.{}.{}".format(n, b)
+            diffusers_unet_map["mid_block.resnets.{}.temporal_res_block.{}".format(i, UNET_MAP_RESNET[b])] = "middle_block.{}.time_stack.{}".format(n, b)
+            #diffusers_unet_map["mid_block.resnets.{}.{}".format(i, UNET_MAP_RESNET[b])] = "middle_block.{}.{}".format(n, b)
+
+    num_res_blocks = list(reversed(num_res_blocks))
+    for x in range(num_blocks):
+        n = (num_res_blocks[x] + 1) * x
+        l = num_res_blocks[x] + 1
+        for i in range(l):
+            c = 0
+            for b in UNET_MAP_RESNET:
+                diffusers_unet_map["up_blocks.{}.resnets.{}.{}".format(x, i, UNET_MAP_RESNET[b])] = "output_blocks.{}.0.{}".format(n, b)
+            c += 1
+            num_transformers = transformer_depth_output.pop()
+            if num_transformers > 0:
+                c += 1
+                for b in UNET_MAP_ATTENTIONS:
+                    diffusers_unet_map["up_blocks.{}.attentions.{}.{}".format(x, i, b)] = "output_blocks.{}.1.{}".format(n, b)
+                for t in range(num_transformers):
+                    for b in TRANSFORMER_BLOCKS:
+                        diffusers_unet_map["up_blocks.{}.attentions.{}.transformer_blocks.{}.{}".format(x, i, t, b)] = "output_blocks.{}.1.transformer_blocks.{}.{}".format(n, t, b)
+            if i == l - 1:
+                for k in ["weight", "bias"]:
+                    diffusers_unet_map["up_blocks.{}.upsamplers.0.conv.{}".format(x, k)] = "output_blocks.{}.{}.conv.{}".format(n, c, k)
+            n += 1
+
+    for k in UNET_MAP_BASIC:
+        diffusers_unet_map[k[1]] = k[0]
+
+    return diffusers_unet_map