Reproduce “SLAT-Phys: Fast Material Property Field Prediction from Structured 3D Latents” 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

Rocktim Jyoti Das, Dinesh Manocha

SLAT-Phys, from the University of Maryland, College Park, introduces an end-to-end network that predicts spatially varying material property fields for 3D objects directly from a single RGB image. The method leverages structured latent representations from a pretrained 3D generative model, achieving a ~120x speedup in inference time compared to prior reconstruction-based approaches while maintaining competitive accuracy.

Key insights

  • Structured latent representations learned by large-scale 3D generative models implicitly encode sufficient physically meaningful information for accurate material property prediction, beyond just geometry and appearance.
  • Directly regressing material properties from these latent features, without explicit multi-view 3D reconstruction, enables significant computational efficiency improvements.
  • A sparse transformer architecture can effectively operate on sparse 3D latent features to predict high-resolution, spatially varying material property fields.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.