ZipSplat: Fewer Gaussians, Better 3D Scene Reconstruction
rsasaki0109 · x · 2026-08-17
ZipSplat is a feed-forward model for 3D Gaussian Splatting that reconstructs scenes from a few unposed images in a single forward pass.
Key Features:
- Eliminates the need for poses, intrinsics, or per-scene optimization.
- Compresses input views into a compact set of scene tokens instead of emitting one Gaussian per pixel.
- Decodes tokens into small groups of Gaussians, achieving state-of-the-art quality with significantly fewer Gaussians.
- Features a single knob to trade off quality for size.
The repository includes inference, evaluation, training code, and pretrained weights.
More from Research
- King's College London uses AI to make string theory testable for the first time — skdh · 2026-08-17
- HEP-TH arXiv Submissions Surge 30% as AI Tools Accelerate Physics Research — skdh · 2026-08-17
- Stanford CS231N 2025 taught by Fei-Fei Li is now free on YouTube — Aiden_Tech_Ai · 2026-08-17
- D-James: Ultra Short Multivariate Signatures Released — jedisct1 · 2026-08-17
- dots3-note: 280B MoE agent learns memory via RL — Aiden_Tech_Ai · 2026-08-17
- Designing LLM Agents to Ask Questions When Facing Uncertainty — aestheticcode · 2026-08-17