DiffGI uses differentiable geometry images to improve thin-shell 3D generation

clovir21 · hf · 2026-07-21

# DiffGI: a differentiable geometry-image pipeline for thin-shell 3D generation The paper argues that most 3D generative models still struggle with thin-shell and non-manifold shapes such as garments because implicit volumetric representations tend to enforce watertight topology. Key ideas: - Replace binary geometry images with a continuous 2D TSDF to preserve boundary positions at subpixel precision. - Use a differentiable Marching Squares procedure so surface losses can backpropagate into the 2D latent space. - Train a compact `32x32` DiffGI-VAE with a geometry-aware normal rendering loss. - Build a transformer-based latent diffusion model with a flow-matching objective on top of that latent space for conditional 3D generation. Experiments on garment and object datasets show better reconstruction fidelity and boundary precision than prior geometry-image and voxel-based methods, while using much less compute.

Original post →

More from Multimodal

Multimodal channel →