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

SILSA. Sliding-Window Slice Latents for Topology-Preserving High-Resolution 3D Generation. By Tianjiao Yu and six others, from the University of Illinois Urbana-Champaign. Published on the first of October. No code repository is linked to it.

High-resolution 3D generation increasingly runs on voxel latents and multi-stage pipelines. A first stage predicts which structure is active, and a later one synthesizes the local geometry. That design works, but it has costs. Continuous surfaces get fragmented into many local tokens, generation becomes expensive, and topological consistency often weakens for shapes that are thin or highly connected, exactly the shapes where a broken link is most visible. A thin strut split across many separately generated tokens has many places where it can come apart.

SILSA represents a shape differently: as compact sliding-window slice latents. Rather than generating costly voxel tokens, it uses a fixed set of overlapping slices along the three canonical axes. Each token summarizes a local window of depth. Because neighbouring windows overlap, the cross-sectional continuity of the surface is preserved, and the whole shape can be generated in a single stage with rectified flow. Where the voxel pipeline needed one pass to decide where structure is and another to fill it in, slices let one pass do the whole job.

A Slice VAE encodes oriented surface samples into slice latents along multiple axes, and reconstructs them through a sparse volumetric decoder. The slices along different directions are coordinated by a Volumetric Anchor Lattice, which gives the directional streams a shared 3D workspace. Topology gets its own supervision at the level of slices: it matches persistence diagrams, and aligns Betti transitions between neighbouring slices, so that structural correctness is trained for directly. In plain terms, the model is told not only what each slice should look like, but how the holes and connected pieces in one slice should carry over into the next.

Against the strongest baseline, SILSA improves PSNR by eight point seven percent, coverage by five point nine six absolute points, and Betti error by nine point two percent. The last of those is the one the topology supervision was aimed at.

The savings are larger still. SILSA uses seventy percent fewer tokens than the next-most compact baseline, and over ninety-eight percent fewer than sparse or hierarchical tokenizers. Training memory drops by forty point four percent, and inference time by fifty-eight point five percent. Qualitatively, thin structures, repeated components and long-range connectivity are better preserved.

The voxel pipeline paid twice, in cost and in topology, and SILSA answers both with one change of representation. A fixed set of overlapping slices along three axes stands in for a mass of voxel tokens, and topology is checked slice by slice, so a smaller budget and fewer broken structures arrive together.