theluyuan/DiffSynth-Studio
1
0
Fork 0
mirror of https://github.com/modelscope/DiffSynth-Studio.git synced 2026-03-18 22:08:13 +00:00
Code Issues Packages Projects Releases Wiki Activity

add wan2.2-fun-A14B inp, control and control-camera (#839)

Browse Source 
* update wan2.2-fun

* update wan2.2-fun

* update wan2.2-fun

* add examples

* update wan2.2-fun

* update wan2.2-fun

* Rename Wan2.2-Fun-A14B-Inp.py to Wan2.2-Fun-A14B-InP.py

---------

Co-authored-by: lzw478614@alibaba-inc.com <lzw478614@alibaba-inc.com>
This commit is contained in:
lzws
2025-08-22 14:20:31 +08:00
committed by GitHub
parent 6a45815b23
commit c795e35142
7 changed files with 183 additions and 10 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
diffsynth/configs/model_config.py
View File

@@ -150,6 +150,8 @@ model_loader_configs = [
(None, "b61c605c2adbd23124d152ed28e049ae", ["wan_video_dit"], [WanModel], "civitai"),
(None, "1f5ab7703c6fc803fdded85ff040c316", ["wan_video_dit"], [WanModel], "civitai"),
(None, "5b013604280dd715f8457c6ed6d6a626", ["wan_video_dit"], [WanModel], "civitai"),
(None, "2267d489f0ceb9f21836532952852ee5", ["wan_video_dit"], [WanModel], "civitai"),
(None, "47dbeab5e560db3180adf51dc0232fb1", ["wan_video_dit"], [WanModel], "civitai"),
(None, "a61453409b67cd3246cf0c3bebad47ba", ["wan_video_dit", "wan_video_vace"], [WanModel, VaceWanModel], "civitai"),
(None, "7a513e1f257a861512b1afd387a8ecd9", ["wan_video_dit", "wan_video_vace"], [WanModel, VaceWanModel], "civitai"),
(None, "cb104773c6c2cb6df4f9529ad5c60d0b", ["wan_video_dit"], [WanModel], "diffusers"),

6
diffsynth/models/wan_video_camera_controller.py
View File

@@ -182,7 +182,7 @@ def process_pose_file(cam_params, width=672, height=384, original_pose_width=128
def generate_camera_coordinates(
direction: Literal["Left", "Right", "Up", "Down", "LeftUp", "LeftDown", "RightUp", "RightDown"],
direction: Literal["Left", "Right", "Up", "Down", "LeftUp", "LeftDown", "RightUp", "RightDown", "In", "Out"],
length: int,
speed: float = 1/54,
origin=(0, 0.532139961, 0.946026558, 0.5, 0.5, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0)
@@ -198,5 +198,9 @@ def generate_camera_coordinates(
coor[13] += speed
if "Down" in direction:
coor[13] -= speed
if "In" in direction:
coor[18] -= speed
if "Out" in direction:
coor[18] += speed
coordinates.append(coor)
return coordinates

37
diffsynth/models/wan_video_dit.py
View File

@@ -294,6 +294,7 @@ class WanModel(torch.nn.Module):
):
super().__init__()
self.dim = dim
self.in_dim = in_dim
self.freq_dim = freq_dim
self.has_image_input = has_image_input
self.patch_size = patch_size
@@ -713,6 +714,42 @@ class WanModelStateDictConverter:
"eps": 1e-6,
"require_clip_embedding": False,
}
elif hash_state_dict_keys(state_dict) == "2267d489f0ceb9f21836532952852ee5":
# Wan2.2-Fun-A14B-Control
config = {
"has_image_input": False,
"patch_size": [1, 2, 2],
"in_dim": 52,
"dim": 5120,
"ffn_dim": 13824,
"freq_dim": 256,
"text_dim": 4096,
"out_dim": 16,
"num_heads": 40,
"num_layers": 40,
"eps": 1e-6,
"has_ref_conv": True,
"require_clip_embedding": False,
}
elif hash_state_dict_keys(state_dict) == "47dbeab5e560db3180adf51dc0232fb1":
# Wan2.2-Fun-A14B-Control-Camera
config = {
"has_image_input": False,
"patch_size": [1, 2, 2],
"in_dim": 36,
"dim": 5120,
"ffn_dim": 13824,
"freq_dim": 256,
"text_dim": 4096,
"out_dim": 16,
"num_heads": 40,
"num_layers": 40,
"eps": 1e-6,
"has_ref_conv": False,
"add_control_adapter": True,
"in_dim_control_adapter": 24,
"require_clip_embedding": False,
}
else:
config = {}
return state_dict, config

35
diffsynth/pipelines/wan_video_new.py
View File

@@ -663,22 +663,23 @@ class WanVideoUnit_ImageEmbedderFused(PipelineUnit):
class WanVideoUnit_FunControl(PipelineUnit):
def __init__(self):
super().__init__(
input_params=("control_video", "num_frames", "height", "width", "tiled", "tile_size", "tile_stride", "clip_feature", "y"),
input_params=("control_video", "num_frames", "height", "width", "tiled", "tile_size", "tile_stride", "clip_feature", "y", "latents"),
onload_model_names=("vae",)
)
def process(self, pipe: WanVideoPipeline, control_video, num_frames, height, width, tiled, tile_size, tile_stride, clip_feature, y):
def process(self, pipe: WanVideoPipeline, control_video, num_frames, height, width, tiled, tile_size, tile_stride, clip_feature, y, latents):
if control_video is None:
return {}
pipe.load_models_to_device(self.onload_model_names)
control_video = pipe.preprocess_video(control_video)
control_latents = pipe.vae.encode(control_video, device=pipe.device, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride).to(dtype=pipe.torch_dtype, device=pipe.device)
control_latents = control_latents.to(dtype=pipe.torch_dtype, device=pipe.device)
y_dim = pipe.dit.in_dim-control_latents.shape[1]-latents.shape[1]
if clip_feature is None or y is None:
clip_feature = torch.zeros((1, 257, 1280), dtype=pipe.torch_dtype, device=pipe.device)
y = torch.zeros((1, 16, (num_frames - 1) // 4 + 1, height//8, width//8), dtype=pipe.torch_dtype, device=pipe.device)
y = torch.zeros((1, y_dim, (num_frames - 1) // 4 + 1, height//8, width//8), dtype=pipe.torch_dtype, device=pipe.device)
else:
y = y[:, -16:]
y = y[:, -y_dim:]
y = torch.concat([control_latents, y], dim=1)
return {"clip_feature": clip_feature, "y": y}
@@ -698,6 +699,8 @@ class WanVideoUnit_FunReference(PipelineUnit):
reference_image = reference_image.resize((width, height))
reference_latents = pipe.preprocess_video([reference_image])
reference_latents = pipe.vae.encode(reference_latents, device=pipe.device)
if pipe.image_encoder is None:
return {"reference_latents": reference_latents}
clip_feature = pipe.preprocess_image(reference_image)
clip_feature = pipe.image_encoder.encode_image([clip_feature])
return {"reference_latents": reference_latents, "clip_feature": clip_feature}
@@ -707,13 +710,14 @@ class WanVideoUnit_FunReference(PipelineUnit):
class WanVideoUnit_FunCameraControl(PipelineUnit):
def __init__(self):
super().__init__(
input_params=("height", "width", "num_frames", "camera_control_direction", "camera_control_speed", "camera_control_origin", "latents", "input_image"),
input_params=("height", "width", "num_frames", "camera_control_direction", "camera_control_speed", "camera_control_origin", "latents", "input_image", "tiled", "tile_size", "tile_stride"),
onload_model_names=("vae",)
)
def process(self, pipe: WanVideoPipeline, height, width, num_frames, camera_control_direction, camera_control_speed, camera_control_origin, latents, input_image):
def process(self, pipe: WanVideoPipeline, height, width, num_frames, camera_control_direction, camera_control_speed, camera_control_origin, latents, input_image, tiled, tile_size, tile_stride):
if camera_control_direction is None:
return {}
pipe.load_models_to_device(self.onload_model_names)
camera_control_plucker_embedding = pipe.dit.control_adapter.process_camera_coordinates(
camera_control_direction, num_frames, height, width, camera_control_speed, camera_control_origin)
@@ -728,14 +732,27 @@ class WanVideoUnit_FunCameraControl(PipelineUnit):
control_camera_latents = control_camera_latents.contiguous().view(b, f // 4, 4, c, h, w).transpose(2, 3)
control_camera_latents = control_camera_latents.contiguous().view(b, f // 4, c * 4, h, w).transpose(1, 2)
control_camera_latents_input = control_camera_latents.to(device=pipe.device, dtype=pipe.torch_dtype)
input_image = input_image.resize((width, height))
input_latents = pipe.preprocess_video([input_image])
pipe.load_models_to_device(self.onload_model_names)
input_latents = pipe.vae.encode(input_latents, device=pipe.device)
y = torch.zeros_like(latents).to(pipe.device)
y[:, :, :1] = input_latents
y = y.to(dtype=pipe.torch_dtype, device=pipe.device)
if y.shape[1] != pipe.dit.in_dim - latents.shape[1]:
image = pipe.preprocess_image(input_image.resize((width, height))).to(pipe.device)
vae_input = torch.concat([image.transpose(0, 1), torch.zeros(3, num_frames-1, height, width).to(image.device)], dim=1)
y = pipe.vae.encode([vae_input.to(dtype=pipe.torch_dtype, device=pipe.device)], device=pipe.device, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)[0]
y = y.to(dtype=pipe.torch_dtype, device=pipe.device)
msk = torch.ones(1, num_frames, height//8, width//8, device=pipe.device)
msk[:, 1:] = 0
msk = torch.concat([torch.repeat_interleave(msk[:, 0:1], repeats=4, dim=1), msk[:, 1:]], dim=1)
msk = msk.view(1, msk.shape[1] // 4, 4, height//8, width//8)
msk = msk.transpose(1, 2)[0]
y = torch.cat([msk,y])
y = y.unsqueeze(0)
y = y.to(dtype=pipe.torch_dtype, device=pipe.device)
return {"control_camera_latents_input": control_camera_latents_input, "y": y}
@@ -1048,7 +1065,7 @@ def model_fn_wan_video(
if clip_feature is not None and dit.require_clip_embedding:
clip_embdding = dit.img_emb(clip_feature)
context = torch.cat([clip_embdding, context], dim=1)
# Add camera control
x, (f, h, w) = dit.patchify(x, control_camera_latents_input)