Research map / 3D vision / Face and shape
SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SILSA, an image-to-3D generator that represents shapes as 384 overlapping sliding-window slice latents across three axes, with a topology-supervised SliceVAE and a shared Volumetric Anchor Lattice for single-stage rectified-flow generation.
Problem
High-resolution 3D generation with voxel latents and multi-stage pipelines fragments surfaces into many tokens, inflates cost, and weakens topological consistency for thin or highly connected shapes.
Approach
Represent shapes with a fixed set of overlapping axis-aligned slice latents (each summarizing a local depth window), coordinate the three axis streams through a shared 3D anchor grid during denoising, and supervise decoded cross-sections with persistence-diagram matching and Betti-transition alignment.
Key results
- On image-to-3D, SILSA reaches PSNR 32.74 vs 30.12 for SparseFlex (8.7% relative gain) and COV 79.08% vs 73.12%.
- SliceVAE reduces Betti-Err from 1.743 (SparseFlex) to 1.582 (9.2% relative reduction) while using 384 tokens.
- Compared with Dora, training memory drops from 14.6GB to 8.7GB and inference from 0.82 to 0.34 s/shape.
Surprising findings
- Direct cross-attention between axes nearly matches VAL on IoU (93.18 vs 93.01), with VAL winning mainly on CD and Betti error.
- Doubling or quadrupling slice count beyond N=128 gives only marginal CD/IoU gains while increasing tokens.
Limitations
- Performance depends on diversity and scale of training distribution; out-of-distribution structures may be reconstructed poorly.
- Output quality depends on input image quality; ambiguous views can yield less faithful 3D structure.
Experimental setup
Datasets: Trellis-500K Toys4K DeepFashion3D
Benchmarks: Toys4K DeepFashion3D
Metrics: CLIP similarity FD KD PSNR LPIPS COV MMD Chamfer Distance F-score IoU
Baselines: Shap-E LN3Diff Direct3D 3DTopia-XL InstantMesh GaussianAnything XCube Dora SAR3D Trellis SparseFlex
Builds on: Trellis SparseFlex PointNet rectified flow DINOv2
Closest papers by meaning
- Native and Compact Structured Latents for 3D Generation (CVPR 2026 · similarity 0.89)
- CraftsMan3D: High-fidelity Mesh Generation with 3D Native Diffusion and Interactive Geometry Refiner (CVPR 2025 · similarity 0.88)
- TopoMesh: High-Fidelity Mesh Autoencoding via Topological Unification (CVPR 2026 · similarity 0.88)
- PartCrafter: Structured 3D Mesh Generation via Compositional Latent Diffusion Transformers (NeurIPS 2025 · similarity 0.88)
- Structured 3D Latents for Scalable and Versatile 3D Generation (CVPR 2025 · similarity 0.87)
- GALA: Geometry-Aware Local Adaptive Grids for Detailed 3D Generation (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.