SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation
SILSA generates high-resolution 3D shapes from sliding-window slice latents, using far fewer tokens and improving topology.
SILSA represents 3D shapes as a fixed set of overlapping slice latents along three axes instead of expensive voxel tokens, enabling single-stage rectified-flow generation. A Slice VAE and Volumetric Anchor Lattice preserve cross-sectional continuity, with persistence-diagram and Betti-transition supervision for topology. Against the strongest baseline it improves PSNR by 8.7%, coverage by 5.96 points, and Betti error by 9.2%, while using 70% fewer tokens than the next-most compact baseline, 40.4% less training memory, and 58.5% less inference time.
- Overlapping multi-axis slice latents replace fragmented high-resolution voxel tokens.
- Slice-level supervision matches persistence diagrams and neighboring Betti transitions.
- PSNR rises 8.7% and Betti error falls 9.2% versus the strongest baseline.
- SILSA uses 70% fewer tokens than the next-most compact baseline.
From the original article
1 min read at arxiv.orgHigh-resolution 3D generation increasingly relies on voxel latents and multi-stage pipelines that first predict active structure and then synthesize local geometry. While effective, this design fragments continuous surfaces into many local tokens, inflates generation cost, and often weakens topological consistency for thin or highly connected shapes.
By Tianjiao Yu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.