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

Tianjiao Yu, Xinzhuo Li, Yifan Shen, Ying Shen, Onkar Susladkar, Ismini Lourentzou
PLAN LabUniversity of Illinois Urbana-Champaign

NeurIPS 2026

Paper Code (Coming soon) Results BibTeX
TL;DRSILSA generates high-resolution 3D shapes from a single image using 384 sliding-window slice tokens, preserving thin structures, holes, repeated parts, and long-range connectivity with single-stage rectified flow.
SILSA-generated detailed 3D shapes, including articulated robots and a creature with thin spines; enlarged insets show small structural details.
384 tokens. Coherent 3D structure. Thin geometry, repeated parts, and long-range connectivity.x-axisy-axisz-axis
384
fixed slice tokens
128 along each of three axes
70.0%
fewer latent tokens
vs. Dora, the next-most compact baseline
58.5%
less inference time
0.34 vs. 0.82 seconds per shape, vs. Dora
9.2%
lower Betti error
vs. SparseFlex in VAE reconstruction

Overview

Sliding-window slice representation

Voxel tokenization divides a surface into many local pieces. SILSA instead summarizes overlapping depth windows along three axes. This compact, fixed layout preserves cross-sectional continuity and supports a single-stage generator.

128 x + 128 y + 128 z = 384 tokens
01 / REPRESENT

Sliding-window SliceVAE

Each token aggregates oriented surface samples across a local depth window. A sparse volumetric decoder reconstructs the high-resolution mesh.

02 / COORDINATE

Volumetric Anchor Lattice

A shared 3D workspace lets the x-, y-, and z-axis streams exchange spatial evidence as the rectified-flow transformer generates the slice latents.

03 / PRESERVE

Slice-wise topology loss

Persistent-homology supervision preserves components and holes. Betti-transition matching aligns where they appear, disappear, split, or merge.

Method

SILSA architecture

A topology-aware SliceVAE learns the representation. An image-conditioned rectified-flow transformer generates its 384 slice latents, and the frozen decoder reconstructs geometry through sparse volumetric upsampling.

From slices to surfaces. Overlapping windows encode local geometry; volumetric anchors connect the axes.Click figure to enlarge

Topology is more than surface similarity.

Thin supports can disconnect, holes can close, and nearby parts can merge even when a reconstruction looks locally similar. SILSA supervises both the topology of each cross-section and the transitions between neighboring slices.

Components + holes + where they change

Qualitative evaluation

VAE reconstruction comparisons

Compare SliceVAE with representative 3D tokenizers on thin supports, dense branches, holes, and repeated structures.

Ground truthSILSASparseFlexTrellisXCubeDora
Surface errorLowHigh
Geometry, normals, and surface error. Normal maps appear in the upper insets; error maps appear in the lower insets.Click figure to enlarge
More reconstruction examples

Quantitative evaluation

Quantitative results

Trained on Trellis-500K. Evaluated on 200 randomly sampled Toys4K assets and 50 in-the-wild images, with no training-set overlap.

Efficiency comparison from the paper. Lower is better for all columns.
MethodTokens ↓Stages ↓Parameters ↓Train memory ↓Train time ↓Inference ↓CD ↓
Dora1,2801124 M14.6 GB0.38 s0.82 s9.76
XCube64,821287 M68.4 GB1.18 s2.74 s3.21
Trellis19,847 ± 4,3122347 M42.7 GB0.71 s1.93 s1.12
SparseFlex87,453 ± 18,2642213 M55.4 GB1.47 s1.15 s0.61
SILSA384196 M8.7 GB0.21 s0.34 s0.59

Training memory and time: batch size 4 on one A100; training time is seconds per iteration. Inference is end-to-end seconds per shape. Variable-length token counts are mean ± standard deviation. ↓ Lower is better.

+8.7% PSNR

32.74 vs. 30.12 for SparseFlex in image-to-3D generation.

+5.96 points coverage

79.08% vs. 73.12% for SparseFlex, an absolute gain.

40.4% less training memory

8.7 GB vs. 14.6 GB for Dora under the reported A100 setup.

The paper

Abstract

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-axis slice latents and reconstructs them with a sparse volumetric decoder, while a Volumetric Anchor Lattice coordinates directional slice streams through a shared 3D workspace. To preserve structural correctness, we introduce slice-level topology supervision that matches persistence diagrams and aligns Betti transitions across neighboring slices.

Experiments show that SILSA improves structural fidelity while substantially reducing generation cost. SILSA improves PSNR by 8.7%, coverage by 5.96 absolute points, and Betti error by 9.2% over the strongest baseline, while using 70.0% fewer tokens than the next-most compact baseline and over 98% fewer tokens than sparse or hierarchical tokenizers, effectively reducing training memory by 40.4% and inference time by 58.5%. Qualitative results further show improved preservation of thin structures, repeated components, and long-range connectivity.

Citation

@inproceedings{yu2026silsa,
  title = {SILSA: Sliding-Window Slice Latents for Topology-Preserving
           High-Resolution 3D Generation},
  author = {Yu, Tianjiao and Li, Xinzhuo and Shen, Yifan and Shen, Ying
            and Susladkar, Onkar and Lourentzou, Ismini},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2026}
}

Accepted at NeurIPS 2026. Paper and code links will be added when released.

Figure