SILSA: Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation
SILSA tackles the cost and topological fragility of high-resolution 3D generation by replacing dense or sparse voxel tokens with compact sliding-window slice latents along the three canonical axes. Each of the 384 fixed tokens summarizes a local depth window, preserving cross-sectional continuity while enabling a single-stage rectified-flow generator. Topology-aware slice supervision and a Volumetric Anchor Lattice then maintain geometric and structural coherence. The paper reports large efficiency gains—over 98% fewer tokens than sparse voxelizers—and solid improvements in PSNR, coverage, and Betti error.
SILSA is a well-motivated and empirically strong contribution. The slice-latent representation elegantly sidesteps data-dependent token inflation, the Volumetric Anchor Lattice provides a principled mechanism for cross-axis coherence, and the topology-aware losses yield measurable gains. However, several headline numbers compare efficiency against Dora (the next-most-compact baseline with 1,280 tokens) rather than the strongest quality competitor SparseFlex, and the evaluation corpus is small—only 200 Toys4K assets plus 50 in-the-wild images. The topology gains are real but modest in absolute terms, with a residual Betti-Err of 1.58.
The fixed 384-token budget is the paper’s most compelling practical advance: it removes multi-stage active-voxel prediction and slashes training memory and inference time. The topology supervision is theoretically well-grounded, and the ablation shows that removing the full $\mathcal{L}_{\mathrm{topo}}$ causes Betti-Err to jump from 1.58 to 4.43. The VAL ablation is equally decisive: without cross-axis communication, CD rises to 5.87, confirming that the shared volumetric workspace is essential rather than decorative.
The main text provides little sensitivity analysis for the topology loss, which relies on a hard occupancy threshold and a straight-through estimator for discrete Betti numbers—choices that can be unstable. The decoder still upsamples to $256^3$ via sparse volumetric stages, so the reported 58.5% inference speedup is measured against Dora and may not hold if decoder extraction dominates wall-clock time. In addition, the 8.7% relative PSNR gain over SparseFlex is only about 2.6 dB, and the evaluation set of 250 shapes is small for modern generative benchmarks, raising questions about variance and generalization. Finally, the "topology-preserving" framing is aspirational: a residual Betti-Err of 1.58 means one to two topological errors per shape persist on average.
The quantitative evidence in Tables 1 and 2 broadly supports the claim that slice latents improve structural fidelity and efficiency. SILSA leads on FD, PSNR, IoU, and Betti-Err against strong open-source baselines trained on the same Trellis-500K data, and the authors transparently note that SparseFlex attains a marginally higher CLIP score. Still, comparisons are partially confounded by differing model scales—SILSA uses 96M parameters versus SparseFlex’s 213M—and by the architectural switch from variable sparse tokens to fixed dense slices. The persistence-diagram matching is computed on sampled cross-sections rather than the full volume, which is a necessary efficiency trade-off but still a proxy for true 3D topology.
Reproducibility is limited because the main text delegates all hyperparameters and training details to Appendix C, which is not provided in the excerpt. Key missing information includes the rectified-flow learning rate, batch size, sampling steps, transformer depth $L$, VAL channels $C$, topology-loss weights $\lambda_{\mathrm{PH}}$ and $\lambda_{\mathrm{trans}}$, the SDF-to-occupancy sharpness $\kappa$, and the number of topology-supervised slices $N_s$. The authors mention a frozen DINOv2 encoder but do not specify the variant, and no code repository or exact data split is cited. Independent reproduction would therefore be difficult without the appendices or supplementary code.
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.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.