River-Zhang / SIFU

[CVPR 2024 Highlight] Official repository for paper "SIFU: Side-view Conditioned Implicit Function for Real-world Usable Clothed Human Reconstruction"
https://river-zhang.github.io/SIFU-projectpage/
MIT License
206 stars 9 forks source link

The difference in back texture between the pre-trained model and my retrained model #45

Open glorioushonor opened 2 weeks ago

glorioushonor commented 2 weeks ago

Hello, Dr. Zhang,

I’ve encountered a strange issue. When I use your provided pre-trained model, I can obtain the correct texture for the back. However, when I use my retrained model, I cannot get the correct back texture. I believe the issue lies with the texture estimation MLP in my trained model, which seems to be estimating incorrect colors. By the way, the geometric metrics of my retrained model during testing are almost identical to those mentioned in the paper. The issue seems to be only with the texture generation MLP.

I generated the ICON and PIFU datasets correctly using the provided scripts. Aside from modifying the dataset access path, I haven’t made any other changes to the code.

image image

Could you provide any insights or suggestions on what might be causing this?

Thank you for your time!