Poor inference quality on non-official Expert-Appearance cases

[GHOST-LOVE-YOU]

Hi, thank you for releasing the MACE-Dance code and checkpoints.

I have been testing the Expert-Appearance inference pipeline. The official provided cases work very well on my side, and the generated results look clear and stable.

0002.mp4

However, when I try to run inference on cases outside the official examples, the results become much worse. I observe obvious snow/noise artifacts, and the person in the generated video becomes blurry and unclear.

I would like to ask whether anyone has tested Expert-Appearance on cases other than the official provided cases, and whether there are any recommended data preparation details or constraints that are important for good inference quality.

For one failed case, I used the following inputs:

• source video:

video.mp4

• Face video:
  https://github.com/user-attachments/assets/1982e082-250c-4df7-b617-2448c77bfd82

• Pose video:

pose_video1.mp4

The inference result contains strong noise artifacts and poor identity/appearance clarity:

• Generated result:

00000.mp4

NOTE: The pose video above was not extracted from the same original video. However, I also tried extracting pose directly from the original video using DWPose, and the inference quality was still very poor.

So I suspect the pose source may not be the main reason for the failure.

• Inference result using the DWPose-extracted pose video:

0000.mp4

More Failed Cases

I will also upload several other failed cases to show that this issue is not an isolated example.

Case 1:

pred_stage2.mp4

Case 2:

pred_stage21.mp4

Case 3:

pred_stage22.mp4

Is this behavior expected when using custom/non-official cases?

Are there any specific requirements for the input video, face video, pose video, resolution, crop, person scale, background, or preprocessing steps that are necessary to reproduce the good quality shown in the official cases?

Any suggestions would be greatly appreciated. Thanks!

[Sun-Happy-YKX]

Hi, thanks for your feedback, and sorry for the inconvenience.

This issue is likely caused by a mismatch or loading problem with the adapter files. When the code was released, I had already left AMAP and no longer had access to the original internal environment. The open-sourcing process was therefore assisted by my colleagues, so there may be some bugs or missing details in the released version.

I really appreciate you reporting this. Due to heavy paper submission and work deadlines recently, I may not be able to reproduce and fix the code immediately. However, I plan to re-release a clean and usable version within the next 3–4 months.

[Sun-Happy-YKX]

@GHOST-LOVE-YOU
Sorry for the delay. This part was previously handled by my mentor, Jiashu Zhu. I have since left the team, and the group is currently occupied with an urgent project on world models, so they have not yet had time to trace back the corresponding checkpoint. Once the current project is completed, we will open-source the correct version as soon as possible. Please stay tuned!