SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation
SILSA replaces voxel-heavy 3D generation with overlapping slice latents meant to keep connected geometry intact.
The paper says the method encodes shapes as fixed sliding-window slices across the three canonical axes, then uses a single-stage rectified-flow generator. Its topology supervision matches persistence diagrams and Betti changes between neighboring slices. In experiments, SILSA reports better PSNR, coverage, and Betti error than the strongest baseline while using far fewer tokens. The authors also claim lower training memory and faster inference, with qualitative gains on thin structures and long-range connectivity. HF Daily Papers' note
The paper says the method encodes shapes as fixed sliding-window slices across the three canonical axes, then uses a single-stage rectified-flow generator. Its topology supervision matches persistence diagrams and Betti changes between neighboring slices. In experiments, SILSA reports better PSNR, coverage, and Betti error than the strongest baseline while using far fewer tokens. The authors also claim lower training memory and faster inference, with qualitative gains on thin structures and long-range connectivity. HF Daily Papers' note
score 4