InfiniSplat: Implicit Gaussian Decoding for Large-Baseline Monocular View Synthesis
InfiniSplat replaces fixed pixel-grid Gaussians with surface-aligned supports for wider viewpoint changes.
The paper says current single-image 3D Gaussian Splatting methods can work for nearby views but lose coherent structure across large baselines. InfiniSplat samples 2D supports using depth-induced local surface geometry, then uses a query-conditioned implicit decoder to predict Gaussian attributes from image features. The authors report state-of-the-art results against single-image feed-forward baselines across multiple cross-dataset novel-view-synthesis evaluations. They also claim zero-shot generalization from Hypersim indoor synthetic training to complex open-world scenes. HF Daily Papers' note
The paper says current single-image 3D Gaussian Splatting methods can work for nearby views but lose coherent structure across large baselines. InfiniSplat samples 2D supports using depth-induced local surface geometry, then uses a query-conditioned implicit decoder to predict Gaussian attributes from image features. The authors report state-of-the-art results against single-image feed-forward baselines across multiple cross-dataset novel-view-synthesis evaluations. They also claim zero-shot generalization from Hypersim indoor synthetic training to complex open-world scenes. HF Daily Papers' note
score 5