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

TL;DR AI
2 min readKey summary
InfiniSplat is a new feed-forward 3D Gaussian splatting method for monocular view synthesis from a single image.
It replaces pixel-aligned primitives with a surface-aligned representation using depth-guided support points and a query-conditioned implicit decoder.
The approach reportedly beats prior single-image baselines on multiple cross-dataset novel view synthesis benchmarks.
It also shows zero-shot generalization from synthetic indoor training to more complex open-world scenes, improving rendering stability under large viewpoint changes.
