DiffGI: Differentiable Geometry Images for High-Fidelity Thin-Shell 3D Generation
The paper’s core move is making geometry-image reconstruction differentiable instead of leaving it as post-processing.
DiffGI represents thin-shell surfaces with a continuous 2D TSDF, aiming to avoid the staircase artifacts of binary occupancy maps. It adds a differentiable Marching Squares step so 3D surface losses can train the 2D latent representation directly. The authors report stronger reconstruction fidelity and boundary precision on garment and object datasets, with lower compute than prior geometry-image and voxel methods. HF Daily Papers' note
DiffGI represents thin-shell surfaces with a continuous 2D TSDF, aiming to avoid the staircase artifacts of binary occupancy maps. It adds a differentiable Marching Squares step so 3D surface losses can train the 2D latent representation directly. The authors report stronger reconstruction fidelity and boundary precision on garment and object datasets, with lower compute than prior geometry-image and voxel methods. HF Daily Papers' note
score 4