search

AiuniAI / Unique3D

[NeurIPS 2024] Unique3D: High-Quality and Efficient 3D Mesh Generation from a Single Image
https://wukailu.github.io/Unique3D/
MIT License
3.11k stars 246 forks source link

quality issues #61

Closed KenNavarro730 closed 4 months ago

KenNavarro730 commented 4 months ago

Since nvdiffrast didnt work on my machine I tried to use the old pytorch3d version to do everything. Replacing recon.py, refine.py and project_mesh.py as necessary to make pytorch3d render. However the results I am getting are much worse, even with doing what the other person in this issues thread of making the resolution In project_mesh.py higher (I changed every H,W,resolution variable to 8192 so maybe its the resolution of the input photo I'm using? Checking that next), is this issue likely arising from nvdiffrast somehow producing much better quality 3d renderings? Or is it something else that the online demo is doing I am not aware of?

KenNavarro730 commented 4 months ago

Here example of quality difference

Screenshot 2024-07-08 at 4 03 47 PM Screenshot 2024-07-08 at 4 03 54 PM
wukailu commented 4 months ago

As far as we know, Pytorch3d, for its own reasons, may not render the image correctly in some cases (e.g. when pytorch3d itself raises a warning about max_faces_per_bins). This can lead to miscalculations in the final result. Meanwhile, pytorch3d, despite having similar rendering results, is not identical to the nvdiffrast rendering, especially at the edges.