Reproduce “SUPRA-TEX: 3D Generation with Super-Aligned Texture” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chen Yang, Jiemin Fang, Fanpeng Meng, Zihao Huang, Jichen Pang, Guanjun Wu, Zhikuan Bao, Qi Tian
SUPRA-Tex generates materials from one image by predicting albedo, metallic, and roughness maps, projecting them onto visible geometry, and encoding these partial observations in the material generator’s latent space. A flow transformer completes the surface using the observations as fixed references. On Complex-Tex, shaded foreground PSNR is 20.64 dB versus 19.44 for Pixal3D with ground-truth geometry, a comparison with geometry fixed.
Key insights
- Representing visible evidence directly in material space gives the 3D generator explicit, spatially grounded anchors rather than requiring it to infer both material properties and surface locations from generic image features.
- The method separates reconstruction of observed regions from completion of unseen surfaces; its coordinate-based observations can also accommodate additional views or edited images without retraining the 3D generator.
- Predicted material estimates remain a bottleneck: oracle material observations perform better, and errors in camera or geometry alignment can misplace otherwise useful evidence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.