DiffGI: Differentiable Geometry Images for High-Fidelity Thin-Shell 3D Generation

DiffGI presents an end-to-end 3D-to-2D geometry-image framework for generating thin-shell and non-manifold surfaces such as garments. It replaces discrete binary occupancy maps with a continuous 2D Truncated Signed Distance Function, allowing boundary positions to be represented at subpixel precision within a fixed grid. The paper also introduces differentiable Marching Squares using analytical linear interpolation, enabling surface losses to backpropagate into the 2D latent space. A DiffGI-VAE compresses surfaces into a 32x32 latent space, followed by a transformer-based latent diffusion model trained with flow matching for conditional 3D generation.