nodes/kijai-ComfyUI-WanVideoWrapper
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge branch 'main' into multitalk

Browse Source
This commit is contained in:
kijai
2025-07-04 16:43:53 +03:00
parent 974dd656da 811dd32713
commit ca3c58bb1a
2 changed files with 871 additions and 837 deletions
Download Patch File Download Diff File Expand all files Collapse all files
nodes.py
+57 -32
View File
@@ -2364,6 +2364,8 @@ class WanVideoVACEStartToEndFrame:
"end_image": ("IMAGE",),
"control_images": ("IMAGE",),
"inpaint_mask": ("MASK", {"tooltip": "Inpaint mask to use for the empty frames"}),
"start_index": ("INT", {"default": 0, "min": 0, "max": 10000, "step": 1, "tooltip": "Index to start from"}),
"end_index": ("INT", {"default": -1, "min": -10000, "max": 10000, "step": 1, "tooltip": "Index to end at"}),
},
}
@@ -2373,51 +2375,74 @@ class WanVideoVACEStartToEndFrame:
CATEGORY = "WanVideoWrapper"
DESCRIPTION = "Helper node to create start/end frame batch and masks for VACE"
def process(self, num_frames, empty_frame_level, start_image=None, end_image=None, control_images=None, inpaint_mask=None):
def process(self, num_frames, empty_frame_level, start_image=None, end_image=None, control_images=None, inpaint_mask=None, start_index=0, end_index=-1):
B, H, W, C = start_image.shape if start_image is not None else end_image.shape
device = start_image.device if start_image is not None else end_image.device
# Convert negative end_index to positive
if end_index < 0:
end_index = num_frames + end_index
# Create output batch with empty frames
out_batch = torch.ones((num_frames, H, W, 3), device=device) * empty_frame_level
# Create mask tensor with proper dimensions
masks = torch.ones((num_frames, H, W), device=device)
if control_images is not None:
# Pre-process all images at once to avoid redundant work
if end_image is not None and (end_image.shape[1] != H or end_image.shape[2] != W):
end_image = common_upscale(end_image.movedim(-1, 1), W, H, "lanczos", "disabled").movedim(1, -1)
if control_images is not None and (control_images.shape[1] != H or control_images.shape[2] != W):
control_images = common_upscale(control_images.movedim(-1, 1), W, H, "lanczos", "disabled").movedim(1, -1)
if start_image is not None and end_image is not None:
if start_image.shape != end_image.shape:
end_image = common_upscale(end_image.movedim(-1, 1), W, H, "lanczos", "disabled").movedim(1, -1)
if control_images is None:
empty_frames = torch.ones((num_frames - start_image.shape[0] - end_image.shape[0], H, W, 3), device=device) * empty_frame_level
else:
empty_frames = control_images[start_image.shape[0]:num_frames - end_image.shape[0]]
out_batch = torch.cat([start_image, empty_frames, end_image], dim=0)
masks[0:start_image.shape[0]] = 0
masks[-end_image.shape[0]:] = 0
elif start_image is not None:
if control_images is None:
empty_frames = torch.ones((num_frames - start_image.shape[0], H, W, 3), device=device) * empty_frame_level
else:
empty_frames = control_images[start_image.shape[0]:num_frames]
out_batch = torch.cat([start_image, empty_frames], dim=0)
masks[0:start_image.shape[0]] = 0
elif end_image is not None:
if control_images is None:
empty_frames = torch.ones((num_frames - end_image.shape[0], H, W, 3), device=device) * empty_frame_level
else:
empty_frames = control_images[:num_frames - end_image.shape[0]]
out_batch = torch.cat([empty_frames, end_image], dim=0)
masks[-end_image.shape[0]:] = 0
# Place start image at start_index
if start_image is not None:
frames_to_copy = min(start_image.shape[0], num_frames - start_index)
if frames_to_copy > 0:
out_batch[start_index:start_index + frames_to_copy] = start_image[:frames_to_copy]
masks[start_index:start_index + frames_to_copy] = 0
# Place end image at end_index
if end_image is not None:
# Calculate where to start placing end images
end_start = end_index - end_image.shape[0] + 1
if end_start < 0: # Handle case where end images won't all fit
end_image = end_image[abs(end_start):]
end_start = 0
frames_to_copy = min(end_image.shape[0], num_frames - end_start)
if frames_to_copy > 0:
out_batch[end_start:end_start + frames_to_copy] = end_image[:frames_to_copy]
masks[end_start:end_start + frames_to_copy] = 0
# Apply control images to remaining frames that don't have start or end images
if control_images is not None:
# Create a mask of frames that are still empty (mask == 1)
empty_frames = masks.sum(dim=(1, 2)) > 0.5 * H * W
# Only process frames that are empty and within control_images range
valid_frames = torch.arange(num_frames, device=device)
valid_indices = (empty_frames & (valid_frames < control_images.shape[0]))
if valid_indices.any():
out_batch[valid_indices] = control_images[:num_frames][valid_indices]
# Apply inpaint mask if provided
if inpaint_mask is not None:
inpaint_mask = common_upscale(inpaint_mask.unsqueeze(1), W, H, "nearest-exact", "disabled").squeeze(1).to(device)
# Handle different mask lengths efficiently
if inpaint_mask.shape[0] > num_frames:
inpaint_mask = inpaint_mask[:num_frames]
elif inpaint_mask.shape[0] < num_frames:
inpaint_mask = inpaint_mask.repeat(num_frames // inpaint_mask.shape[0] + 1, 1, 1)[:num_frames]
repeat_factor = (num_frames + inpaint_mask.shape[0] - 1) // inpaint_mask.shape[0] # Ceiling division
inpaint_mask = inpaint_mask.repeat(repeat_factor, 1, 1)[:num_frames]
# Apply mask in one operation
masks = inpaint_mask * masks
empty_mask = torch.ones_like(masks, device=device)
masks = inpaint_mask * empty_mask
return (out_batch.cpu().float(), masks.cpu().float())