How to achieve the effect of subject removal?

[zijieloooooou]

This is source video:

source.mp4

This is mask video:

mask.mp4

This is the command I used:

python vace/vace_wan_inference.py \
             --src_video ./source.mp4 \
             --src_mask ./mask.mp4 \
             --prompt "This video captures a low-angle, upward view into a dense, dark forest. Tall, slender tree trunks rise vertically, their forms silhouetted against a muted, overcast sky that peeks through the thick canopy of leaves. The lighting is dim and atmospheric, creating deep shadows and a somber, moody ambiance. "

These are the generated results:
Why wasn't that person in out_video.mp4 removed?

out_video.mp4

src_mask.mp4

src_video.mp4

[zijieloooooou]

After testing, it was found that the area to be removed in the source video needs to be blacked out first.

src_video.mp4

src_mask.mp4

This is the result video:

out_video.mp4

[zijieloooooou]

The generated results for Wan2.1-VACE-1.3B are normal. However, the generated results for Wan2.1-VACE-14B are abnormal. Their inputs (including prompts) are all the same, only the weight file is different.

Wan2.1-VACE-14B_out_video.mp4

[ruolinsss]

I also found the performance of subject removal is not stable, related to input_video/mask/prompt/model weights, and most of the cases, not good.

[ppc2017]

After testing, it was found that the area to be removed in the source video needs to be blacked out first.

You must put solid gray (rgb 127,127,127) in the source video, not black