SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation
SILSA replaces voxel-token 3D generation with overlapping slice latents meant to keep surfaces connected while cutting cost.
The paper says its fixed multi-axis slices summarize local depth windows, supporting single-stage rectified-flow generation. It adds topology supervision using persistence diagrams and Betti transitions between neighboring slices. Reported gains include 8.7% higher PSNR, 5.96 points higher coverage, and 9.2% lower Betti error versus the strongest baseline. The authors also report 70.0% fewer tokens than the next-most compact baseline, with lower memory use and faster inference. ArXiv · AI/CL/LG's note
The paper says its fixed multi-axis slices summarize local depth windows, supporting single-stage rectified-flow generation. It adds topology supervision using persistence diagrams and Betti transitions between neighboring slices. Reported gains include 8.7% higher PSNR, 5.96 points higher coverage, and 9.2% lower Betti error versus the strongest baseline. The authors also report 70.0% fewer tokens than the next-most compact baseline, with lower memory use and faster inference. ArXiv · AI/CL/LG's note
score 4