UniqueSplat: View-Conditioned 3D Gaussian Splatting for Generalizable Reconstruction
zhenjun_zhao · x · 2026-08-04
UniqueSplat introduces a view-conditioned feed-forward 3D Gaussian Splatting (3DGS) model to address the limitations of existing methods (like pixelSplat and MVSplat) that generate fixed Gaussians across all views without adapting to specific target viewpoints.
Key innovations include:
- View-Conditioned Prior: Learns view-conditioned information as a prior integrated into network parameters, dynamically adjusting Gaussians based on the target view.
- Two-Branch HyperNetwork: Simultaneously learns view-agnostic embeddings and view-specific knowledge, exploring shareable features while adapting to specific views at test time.
Experiments on datasets like RealEstate10K, ACID, and DTU demonstrate that UniqueSplat outperforms state-of-the-art methods and shows strong generalization in cross-dataset evaluations.
More from Research
- Behind NeurIPS Rejections: Limited Physical Space Forces Extra Papers Out — qberthet · 2026-08-04
- Include Standard Libraries in RL Environments to Stop AI Agents Reinventing the Wheel — cephaloform · 2026-08-04
- SimpleAI releases HiFi-UMI: high-fidelity data engine enables direct real-robot deployment without teleoperation data — 机器之心 · 2026-08-04
- Qianxun AI study: legacy data shows 'emergent transfer' after robot hardware upgrades, with a critical threshold — 机器之心 · 2026-08-04
- QuerySplat: Decoupling Geometry and Appearance in 3DGS Prediction — zhenjun_zhao · 2026-08-04
- UniSim-SLAM: Feed-Forward SLAM with Unified Sim(3) Optimization — zhenjun_zhao · 2026-08-04