feat: Support SCAIL WanVideo model

comfy/ldm/wan/model.py

@@ -1621,3 +1621,114 @@ def block_wrap(args):
         # unpatchify
         x = self.unpatchify(x, grid_sizes)
         return x
+
+class SCAILWanModel(WanModel):
+    def __init__(self, model_type="scail", patch_size=(1, 2, 2), in_dim=20, dim=5120, operations=None, device=None, dtype=None, **kwargs):
+        super().__init__(model_type='i2v', patch_size=patch_size, in_dim=in_dim, dim=dim, operations=operations, device=device, dtype=dtype, **kwargs)
+
+        self.patch_embedding_pose = operations.Conv3d(in_dim, dim, kernel_size=patch_size, stride=patch_size, device=device, dtype=torch.float32)
+
+    def forward_orig(self, x, t, context, clip_fea=None, freqs=None, transformer_options={}, pose_latents=None, reference_latent=None, **kwargs):
+
+        if reference_latent is not None:
+            x = torch.cat((reference_latent, x), dim=2)
+
+        # embeddings
+        x = self.patch_embedding(x.float()).to(x.dtype)
+        grid_sizes = x.shape[2:]
+        transformer_options["grid_sizes"] = grid_sizes
+        x = x.flatten(2).transpose(1, 2)
+
+        scail_pose_seq_len = 0
+        if pose_latents is not None:
+            scail_x = self.patch_embedding_pose(pose_latents.float()).to(x.dtype)
+            scail_x = scail_x.flatten(2).transpose(1, 2)
+            scail_pose_seq_len = scail_x.shape[1]
+            x = torch.cat([x, scail_x], dim=1)
+            del scail_x
+
+        # time embeddings
+        e = self.time_embedding(sinusoidal_embedding_1d(self.freq_dim, t.flatten()).to(dtype=x[0].dtype))
+        e = e.reshape(t.shape[0], -1, e.shape[-1])
+        e0 = self.time_projection(e).unflatten(2, (6, self.dim))
+
+        # context
+        context = self.text_embedding(context)
+
+        context_img_len = None
+        if clip_fea is not None:
+            if self.img_emb is not None:
+                context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
+                context = torch.cat([context_clip, context], dim=1)
+            context_img_len = clip_fea.shape[-2]
+
+        patches_replace = transformer_options.get("patches_replace", {})
+        blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.blocks)
+        transformer_options["block_type"] = "double"
+        for i, block in enumerate(self.blocks):
+            transformer_options["block_index"] = i
+            if ("double_block", i) in blocks_replace:
+                def block_wrap(args):
+                    out = {}
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], context_img_len=context_img_len, transformer_options=args["transformer_options"])
+                    return out
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs, "transformer_options": transformer_options}, {"original_block": block_wrap})
+                x = out["img"]
+            else:
+                x = block(x, e=e0, freqs=freqs, context=context, context_img_len=context_img_len, transformer_options=transformer_options)
+
+        # head
+        x = self.head(x, e)
+
+        if scail_pose_seq_len > 0:
+            x = x[:, :-scail_pose_seq_len]
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+
+        if reference_latent is not None:
+            x = x[:, :, reference_latent.shape[2]:]
+
+        return x
+
+    def rope_encode(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None, pose_latents=None, transformer_options={}):
+        main_freqs = super().rope_encode(t, h, w, t_start=t_start, steps_t=steps_t, steps_h=steps_h, steps_w=steps_w, device=device, dtype=dtype, transformer_options=transformer_options)
+
+        if pose_latents is None:
+            return main_freqs
+
+        F_pose, H_pose, W_pose = pose_latents.shape[-3], pose_latents.shape[-2], pose_latents.shape[-1]
+
+        # if pose is at half resolution, scale_y/scale_x=2 stretches the position range to cover the same RoPE extent as the main frames
+        h_scale = h / H_pose
+        w_scale = w / W_pose
+
+        # 120 w-offset and shift 0.5 to place positions at midpoints (0.5, 2.5, ...) to match the original code
+        h_shift = (h_scale - 1) / 2
+        w_shift = (w_scale - 1) / 2
+        pose_transformer_options = {"rope_options": {"shift_y": h_shift, "shift_x": 120.0 + w_shift, "scale_y": h_scale, "scale_x": w_scale}}
+        pose_freqs = super().rope_encode(F_pose, H_pose, W_pose, t_start=t_start, device=device, dtype=dtype, transformer_options=pose_transformer_options)
+
+        return torch.cat([main_freqs, pose_freqs], dim=1)
+
+    def _forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, pose_latents=None, **kwargs):
+        bs, c, t, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
+
+        if pose_latents is not None:
+            pose_latents = comfy.ldm.common_dit.pad_to_patch_size(pose_latents, self.patch_size)
+
+        t_len = t
+        if time_dim_concat is not None:
+            time_dim_concat = comfy.ldm.common_dit.pad_to_patch_size(time_dim_concat, self.patch_size)
+            x = torch.cat([x, time_dim_concat], dim=2)
+            t_len = x.shape[2]
+
+        reference_latent = None
+        if "reference_latent" in kwargs:
+            reference_latent = comfy.ldm.common_dit.pad_to_patch_size(kwargs.pop("reference_latent"), self.patch_size)
+            t_len += reference_latent.shape[2]
+
+        freqs = self.rope_encode(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options, pose_latents=pose_latents)
+        return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs, transformer_options=transformer_options, pose_latents=pose_latents, reference_latent=reference_latent, **kwargs)[:, :, :t, :h, :w]

comfy/model_base.py

@@ -1466,6 +1466,59 @@ def process_timestep(self, timestep, x, denoise_mask=None, **kwargs):
     def scale_latent_inpaint(self, sigma, noise, latent_image, **kwargs):
         return latent_image
 
+class WAN21_SCAIL(WAN21):
+    def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
+        super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.SCAILWanModel)
+        self.memory_usage_factor_conds = ("reference_latent", "pose_latents")
+        self.memory_usage_shape_process = {"pose_latents": lambda shape: [shape[0], shape[1], 1.5, shape[-2], shape[-1]]}
+        self.image_to_video = image_to_video
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+
+        reference_latents = kwargs.get("reference_latents", None)
+        if reference_latents is not None:
+            ref_latent = self.process_latent_in(reference_latents[-1])
+            ref_mask = torch.ones_like(ref_latent[:, :4])
+            ref_latent = torch.cat([ref_latent, ref_mask], dim=1)
+            out['reference_latent'] = comfy.conds.CONDRegular(ref_latent)
+
+        pose_latents = kwargs.get("pose_video_latent", None)
+        if pose_latents is not None:
+            pose_latents = self.process_latent_in(pose_latents)
+            pose_mask = torch.ones_like(pose_latents[:, :4])
+            pose_latents = torch.cat([pose_latents, pose_mask], dim=1)
+            out['pose_latents'] = comfy.conds.CONDRegular(pose_latents)
+
+        pose_start = kwargs.get("pose_start", 0.0)
+        pose_end = kwargs.get("pose_end", 1.0)
+        out['pose_start'] = comfy.conds.CONDConstant(pose_start)
+        out['pose_end'] = comfy.conds.CONDConstant(pose_end)
+
+        return out
+
+    def apply_model(self, x, t, c_concat=None, c_crossattn=None, control=None, transformer_options={}, pose_start=0.0, pose_end=1.0, **kwargs):
+        if t[0] >= self.model_sampling.percent_to_sigma(pose_start) or t[0] <= self.model_sampling.percent_to_sigma(pose_end):
+            kwargs.pop("pose_latents", None)
+
+        return comfy.patcher_extension.WrapperExecutor.new_class_executor(
+            self._apply_model,
+            self,
+            comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.APPLY_MODEL, transformer_options)
+        ).execute(x, t, c_concat, c_crossattn, control, transformer_options, **kwargs)
+
+    def extra_conds_shapes(self, **kwargs):
+        out = {}
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            out['reference_latent'] = list([1, 20, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16])
+
+        pose_latents = kwargs.get("pose_video_latent", None)
+        if pose_latents is not None:
+            out['pose_latents'] = [pose_latents.shape[0], 20, *pose_latents.shape[2:]]
+
+        return out
+
 class Hunyuan3Dv2(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan3d.model.Hunyuan3Dv2)

comfy/model_detection.py

@@ -496,6 +496,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["model_type"] = "humo"
         elif '{}face_adapter.fuser_blocks.0.k_norm.weight'.format(key_prefix) in state_dict_keys:
             dit_config["model_type"] = "animate"
+        elif '{}patch_embedding_pose.weight'.format(key_prefix) in state_dict_keys:
+            dit_config["model_type"] = "scail"
         else:
             if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys:
                 dit_config["model_type"] = "i2v"

comfy/supported_models.py

@@ -1256,6 +1256,16 @@ def get_model(self, state_dict, prefix="", device=None):
         out = model_base.WAN22(self, image_to_video=True, device=device)
         return out
 
+class WAN21_SCAIL(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "scail",
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN21_SCAIL(self, image_to_video=False, device=device)
+        return out
+
 class Hunyuan3Dv2(supported_models_base.BASE):
     unet_config = {
         "image_model": "hunyuan3d2",
@@ -1667,6 +1677,6 @@ def clip_target(self, state_dict={}):
         return supported_models_base.ClipTarget(comfy.text_encoders.ace15.ACE15Tokenizer, comfy.text_encoders.ace15.te(**detect))
 
 
-models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima]
+models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, WAN21_SCAIL, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima]
 
 models += [SVD_img2vid]

comfy_extras/nodes_wan.py

@@ -1456,6 +1456,63 @@ def execute(cls, mode: DCValues, model, model_patch, positive, negative, vae, wi
         return io.NodeOutput(model_patched, positive, negative, out_latent, trim_image)
 
 
+class WanSCAILToVideo(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanSCAILToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=512, min=32, max=nodes.MAX_RESOLUTION, step=32),
+                io.Int.Input("height", default=896, min=32, max=nodes.MAX_RESOLUTION, step=32),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.ClipVisionOutput.Input("clip_vision_output", optional=True),
+                io.Image.Input("reference_image", optional=True),
+                io.Image.Input("pose_video", optional=True, tooltip="Video used for pose conditioning. Will be downscaled to half the resolution of the main video."),
+                io.Float.Input("pose_strength", default=1.0, min=0.0, max=10.0, step=0.01, tooltip="Strength of the pose latent."),
+                io.Float.Input("pose_start", default=0.0, min=0.0, max=1.0, step=0.01, tooltip="Start step to use pose conditioning."),
+                io.Float.Input("pose_end", default=1.0, min=0.0, max=1.0, step=0.01, tooltip="End step to use pose conditioning."),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent", tooltip="Empty latent of the generation size."),
+            ],
+            is_experimental=True,
+        )
+
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, pose_strength, pose_start, pose_end, reference_image=None, clip_vision_output=None, pose_video=None) -> io.NodeOutput:
+        latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
+
+        ref_latent = None
+        if reference_image is not None:
+            reference_image = comfy.utils.common_upscale(reference_image[:1].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+            ref_latent = vae.encode(reference_image[:, :, :, :3])
+
+        if ref_latent is not None:
+            positive = node_helpers.conditioning_set_values(positive, {"reference_latents": [ref_latent]}, append=True)
+            negative = node_helpers.conditioning_set_values(negative, {"reference_latents": [torch.zeros_like(ref_latent)]}, append=True)
+
+        if clip_vision_output is not None:
+            positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output})
+            negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output})
+
+        if pose_video is not None:
+            pose_video = comfy.utils.common_upscale(pose_video[:length].movedim(-1, 1), width // 2, height // 2, "area", "center").movedim(1, -1)
+            pose_video_latent = vae.encode(pose_video[:, :, :, :3]) * pose_strength
+            positive = node_helpers.conditioning_set_values(positive, {"pose_video_latent": pose_video_latent, "pose_start": pose_start, "pose_end": pose_end})
+            negative = node_helpers.conditioning_set_values(negative, {"pose_video_latent": pose_video_latent, "pose_start": pose_start, "pose_end": pose_end})
+
+        out_latent = {}
+        out_latent["samples"] = latent
+        return io.NodeOutput(positive, negative, out_latent)
+
+
 class WanExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[io.ComfyNode]]:
@@ -1476,6 +1533,7 @@ async def get_node_list(self) -> list[type[io.ComfyNode]]:
             WanAnimateToVideo,
             Wan22ImageToVideoLatent,
             WanInfiniteTalkToVideo,
+            WanSCAILToVideo,
         ]
 
 async def comfy_entrypoint() -> WanExtension: