Spatial Sparse Attention (SSA)
Attention tailored to sparse volumetric tokens with reported multi-fold training/inference speedups in the paper.
DreamTechAI · Open image-to-3D
Direct3D-S2
Gigascale sparse-volume 3D generation with Spatial Sparse Attention
Direct3D-S2 is a scalable 3D generation framework on sparse volumes. Its Spatial Sparse Attention (SSA) mechanism speeds Diffusion Transformer compute on sparse 3D tokens, enabling high-resolution sparse SDF generation—described in arXiv:2505.17412 (NeurIPS 2025).
What is Direct3D-S2?
Dense full-volume 3D transformers are expensive. Direct3D-S2 keeps a consistent sparse volumetric representation through VAE encode/decode and generation, avoiding heterogeneous intermediate formats that hurt training stability.
SSA partitions sparse tokens using 3D coordinates and combines compressed global context, blockwise selection, and local windows. Authors report about 3.9× forward and 9.6× backward speedups versus baseline sparse attention patterns, and demonstrate training at 1024³ resolution with only 8 GPUs—workloads they contrast with denser setups needing far more GPUs at lower resolutions.
Pixal3D’s paper-branch lineage historically relates to Direct3D-S2-style sparse generation, making this demo a useful open counterpart when comparing pixel-aligned methods against canonical sparse-voxel generators.
When to use Direct3D-S2
Use Direct3D-S2 when evaluating sparse high-resolution shape generation efficiency and quality—especially if you care about SSA-style DiT scaling and sparse SDF pipelines rather than instant low-poly previews.
How Direct3D-S2 works
STEP 1
Encode to sparse latent volumes
A sparse VAE maps shapes into structured sparse latent tokens instead of dense grids everywhere.
STEP 2
Diffuse with SSA-enabled DiT
An image-conditioned sparse DiT updates tokens efficiently using Spatial Sparse Attention.
STEP 3
Decode high-resolution SDF / mesh
Decode sparse outputs to surfaces suitable for mesh extraction in the demo pipeline.
Key capabilities
Unified sparse VAE
Keeps sparse volumetric format across input, latent, and output stages for stabler high-res training.
Gigascale-oriented design
Authors highlight practical 1024³-scale training with 8 GPUs as a key systems contribution.
Public demo
Hugging Face Space wushuang98/Direct3D-S2-v1.0-demo for interactive trials.
Technical snapshot
Figures below come from the linked papers, model cards, or project pages—not from unbenchmarked third-party marketing claims.
- Paper
- arXiv:2505.17412 (NeurIPS 2025)
- Key mechanism
- Spatial Sparse Attention (SSA) on sparse volumes
- Reported efficiency
- ~3.9× forward / ~9.6× backward speedups (paper)
- Training note
- 1024³ resolution with 8 GPUs (authors’ claim)
- Code
- github.com/DreamTechAI/Direct3D-S2
- Demo
- spaces/wushuang98/Direct3D-S2-v1.0-demo
Who it is for
High-resolution shape research
Benchmark sparse 3D generators against TRELLIS / Hunyuan / reconstruction models.
Detail-heavy objects
Sparse occupancy focuses capacity on solid structure instead of empty space.
Systems comparison
Study how SSA-style attention changes feasibility of large 3D token sets.
Limitations & practical notes
- Public Spaces may run on limited hardware versus the paper’s full training/inference setup.
- Speedup and resolution figures are from the paper’s experimental setting—not guaranteed in the embedded demo.
- Textured PBR quality may differ from multi-stage paint systems like Hunyuan3D-2.1.
Frequently asked questions
What is SSA?+
Spatial Sparse Attention is Direct3D-S2’s attention design for sparse 3D tokens—combining compressed global context, selected spatial blocks, and local windows (see project page / paper figures).
Is this the same as Pixal3D?+
No. Pixal3D is a separate pixel-aligned system. Direct3D-S2 is an independent sparse-volume generator that appears in related research comparisons.
Sources & references
Landing claims are grounded in the following primary materials. Always prefer the original paper or model card for citations.
Try Direct3D-S2 in the browser
Jump into the playground and generate a 3D model from your own image—no install required.