3DRAE: Any 3D Scene in 1K Tokens — Diffusion in 3D Latent Space for Scene Generation
Dongxu Wei · hf · 2026-10-21
Researchers from Westlake University, Afari Intelligent Drive, and Zhejiang University present 3DRAE, a new representation for scene-level 3D generation.
- It grounds any number of multi-view images (at any resolution/aspect ratio) into a fixed-length 3D latent — e.g., 1,024 tokens — turning "any 3D scene" into a compact token sequence.
- Unlike 2D diffusion methods that rely on variable-size multi-view 2D latents (e.g., 8,192 tokens) and are view-coupled, direct diffusion in the 3D latent space improves spatial consistency and efficiency.
- A decoder with arbitrary pose queries reconstructs images, point maps, and colored 3D point clouds from novel views.
- Project page: wsdx.github.io/3DRAE
More from Research
- Stanford's FACT model uses human force signals to beat T-Rex by up to 60 points on dexterous tasks — stanfordnlp · 2026-10-21
- Model Nails Ligand Position on Homolog with Known Structure — johnowhitaker · 2026-10-21
- 2721 simulated personas in an artificial city to model how ideas go viral — ycombinator · 2026-10-21
- NeurIPS paper: calibrated probability talk beats action talk for multi-agent collaboration — Aaroth · 2026-10-21
- MIT's Adam Chlipala makes the case for Bluespec-style stepwise derivation over AI codegen — satnam6502 · 2026-10-21
- Kidger's verdict on RN's bio-AI launch: there's fire beneath the smoke — PatrickKidger · 2026-10-21