nodes/stuttlepress-ComfyUI-Wan-VACE-Prep
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Remove reference_image input from WanVACEInpaint

Browse Source 
Handle reference images via WanVaceToVideo.reference_image instead,
which properly VAE-encodes, channels-doubles, and prepends the frame
alongside the control video latent. Avoids double-prepending when
both nodes tried to handle it.
This commit is contained in:
stuttlepress
2026-06-01 13:57:43 -05:00
parent fd26858b55
commit bb8fbf18c0
2 changed files with 7 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+4 -5
View File
@@ -220,7 +220,7 @@ Stuff here is new and has not been thoroughly tested. Inputs, outputs, and behav
> **Experimental** - new and not well tested. Behavior and interface may change.
Prepares control video and mask for inpainting. An optional reference image can be input as a context frame to guide generation.
Prepares control video and mask for inpainting. Connect the output to `WanVaceToVideo.control_video` / `control_masks`; use `WanVaceToVideo.reference_image` separately if you need a reference frame.
**Parameters:**
@@ -228,16 +228,15 @@ Prepares control video and mask for inpainting. An optional reference image can
|-|-|-|
| video | | Source video frames (IMAGE) |
| mask | | Inpaint mask. White (1) marks regions to regenerate, black (0) preserves the original. Can be a single frame (broadcast to all frames) or a per-frame sequence. |
| reference_image | *(optional)* | A single image prepended as a context frame with mask=0. Resized automatically if dimensions differ from the video. |
**Outputs:**
| Output | Description |
|-|-|
| control_video | VACE control video input. Masked pixels replaced with gray (0.5); reference frame prepended if supplied. |
| control_mask | VACE control mask input. Matches control_video length; reference frame slot is fully black (0). |
| control_video | VACE control video input. Masked pixels replaced with gray (0.5). |
| control_mask | VACE control mask input. White (1) where inpainting should occur, black (0) over preserved content. |
| width, height | Video dimensions (must be divisible by 16) |
| length | Frame count (video frames + 1 if reference image provided) |
| length | Frame count (matches input video) |
---
vace_inpaint.py
+3 -29
View File
@@ -1,5 +1,4 @@
import torch
import torch.nn.functional as F
class WanVACEInpaint:
@@ -10,9 +9,6 @@ class WanVACEInpaint:
"video": ("IMAGE",),
"mask": ("MASK",),
},
"optional": {
"reference_image": ("IMAGE",),
},
}
RETURN_TYPES = ("IMAGE", "MASK", "INT", "INT", "INT")
@@ -22,12 +18,11 @@ class WanVACEInpaint:
DESCRIPTION = (
"Prepares a video for VACE inpainting. Masked regions (mask=1) are "
"replaced with a gray placeholder so Wan VACE will regenerate them while "
"preserving the rest. An optional reference image is prepended as a "
"context frame (mask=0) to guide generation."
"preserving the rest."
)
EXPERIMENTAL = True
def run(self, video, mask, reference_image=None):
def run(self, video, mask):
N, H, W, C = video.shape
if W % 16 != 0 or H % 16 != 0:
@@ -58,27 +53,6 @@ class WanVACEInpaint:
mask_bool = mask > 0.5 # [N, H, W]
masked_video[mask_bool] = 0.5
if reference_image is not None:
ref = reference_image[0:1] # [1, H_ref, W_ref, C]
ref_h, ref_w = ref.shape[1], ref.shape[2]
if ref_h != H or ref_w != W:
print(
f"[WanVACEInpaint] Resizing reference image from "
f"{ref_w}x{ref_h} to {W}x{H}"
)
ref = ref.permute(0, 3, 1, 2) # [1, C, H_ref, W_ref]
ref = F.interpolate(ref, size=(H, W), mode="bilinear", align_corners=False)
ref = ref.permute(0, 2, 3, 1) # [1, H, W, C]
control_video = torch.cat([ref, masked_video], dim=0)
ref_mask = torch.zeros(1, H, W, dtype=mask.dtype, device=mask.device)
control_mask = torch.cat([ref_mask, mask], dim=0)
length = N + 1
else:
control_video = masked_video
control_mask = mask
length = N
return (control_video, control_mask, W, H, length)
return (masked_video, mask, W, H, N)