High-resolution 3D generation has increasingly depended on voxel latents and multi-stage pipelines that first predict an active structure and then synthesize local geometry. According to the paper, this approach fragments continuous surfaces into numerous local tokens, raises generation cost, and often degrades topological consistency for thin or highly connected shapes.
To address these issues, the authors introduce SILSA, a topology-aware framework that represents shapes with compact sliding-window slice latents. Rather than generating costly voxel tokens, SILSA uses a fixed set of overlapping slices along the three canonical axes, with each token summarizing a local depth window to preserve cross-sectional continuity and enable single-stage rectified-flow generation.
The method pairs a Slice VAE — which encodes oriented surface samples into multi-axis slice latents and reconstructs them with a sparse volumetric decoder — with a Volumetric Anchor Lattice that coordinates directional slice streams through a shared 3D workspace. For structural correctness, the authors add slice-level topology supervision that matches persistence diagrams and aligns Betti transitions across neighboring slices.
Experiments report that SILSA improves PSNR by 8.7%, coverage by 5.96 absolute points, and Betti error by 9.2% over the strongest baseline. It also uses 70.0% fewer tokens than the next-most compact baseline and over 98% fewer tokens than sparse or hierarchical tokenizers, reducing training memory by 40.4% and inference time by 58.5%.
Qualitative results indicate better preservation of thin structures, repeated components, and long-range connectivity. The paper notes that SILSA generates high-resolution 3D shapes from a single image using 384 sliding-window slice tokens.