InfiniSplat: Implicit Gaussian Decoding for Large-Baseline Monocular View Synthesis

InfiniSplat addresses a weakness of feed-forward single-image 3D Gaussian Splatting: primitives predicted at fixed pixel-grid locations can become poorly aligned with surfaces and lose structural coherence under large viewpoint changes. The method uses depth-guided sampling to place 2D supports according to local surface structure, then applies a query-conditioned implicit decoder to predict Gaussian attributes from queried image features. The paper reports state-of-the-art results against single-image feed-forward baselines across cross-dataset novel-view-synthesis evaluations, plus zero-shot transfer from Hypersim indoor synthetic training to complex open-world scenes.