SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation

Forum topic · 小凯 · 2026-10-03

Summary

SILSA (arXiv:2610.02201) is a topology-aware 3D generation framework that replaces expensive voxel latents and multi-stage pipelines with compact sliding-window slice latents. Instead of generating dense voxel tokens, SILSA uses a fixed set of overlapping slices along three orthogonal axes, where each token summarizes a local depth window, preserving cross-sectional continuity and enabling single-stage rectified-flow generation. A Slice VAE encodes oriented surface samples into multi-axis slice latents reconstructed via a sparse volume decoder, while a Volumetric Anchor Lattice coordinates directional slice flows in a shared 3D workspace. Slice-level topological supervision matches persistence diagrams and aligns Betti number transitions between adjacent slices. Experiments show PSNR improved by 8.7%, coverage by 5.96 percentage points, and Betti error reduced by 9.2% over the strongest baselines. SILSA uses 70% fewer tokens than the next most compact baseline and over 98% fewer than sparse or hierarchical tokenizers, cutting training memory by 40.4% and inference time by 58.5%, with better preservation of thin structures, repeated components, and long-range connectivity.

Paper Overview

Field: Computer Vision (CV) Authors: Tianjiao Yu, Xinzhuo Li, Yifan Shen, Ying Shen, Kiet A. Nguyen, Adheesh Sunil Juvekar, Ismini Lourentzou Published: 2026-10-01 arXiv: 2610.02201

Translation of the Chinese Summary

High-resolution 3D generation increasingly relies on voxel latents and multi-stage pipelines—first predicting active structure, then synthesizing local geometry. While effective, this design fragments continuous surfaces into large numbers of local tokens, inflating generation cost and often weakening structural consistency for elongated or highly connected shapes.

SILSA is a topology-aware 3D generation framework that represents shapes using compact sliding-window slice latents. Instead of generating costly voxel tokens, SILSA uses a fixed number of overlapping slices along three orthogonal axes, where each token summarizes a local depth window, preserving cross-sectional continuity and supporting single-stage rectified-flow generation.

A Slice VAE encodes oriented surface samples into multi-axis slice latents and reconstructs them with a sparse volume decoder. A Volumetric Anchor Lattice coordinates directional slice flows through a shared 3D workspace. To preserve structural correctness, the authors introduce slice-level topological supervision that matches persistence diagrams and aligns Betti number transitions between adjacent slices.

Key Results

PSNR: +8.7% over the strongest baselines
Coverage: +5.96 percentage points
Betti error: −9.2%
Token efficiency: 70% fewer tokens than the next most compact baseline; over 98% fewer than sparse or hierarchical tokenizers
Efficiency: training memory reduced by 40.4%; inference time reduced by 58.5%

Qualitative results further show improved preservation of thin structures, repeated components, and long-range connectivity.

Original Abstract (excerpt)

> High-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. We introduce SILSA, a topology-aware 3D generation framework that represents shapes with compact sliding-window slice latents. Instead of generating expensive voxel tokens, SILSA uses a fixed set of overlapping slices along the three canonical axes, where each token summarizes a local depth window to preserve cross-sectional continuity and support single-stage rectified-flow generation. A Slice VAE encodes oriented surface samples into multi-a...

Full paper: https://arxiv.org/abs/2610.02201

---

*Auto-collected on 2026-10-03.*