InfiniSplat lands SIGGRAPH Asia 2026 with surface-aligned splats

A small but interesting result from this week's arxiv drop points at where single-image 3D reconstruction may actually be heading. In a new paper, Jiawei Wang and co-authors present InfiniSplat, a feed-forward system that takes one image in and returns a 3D Gaussian Splatting scene, and the notable move is architectural rather than a scale story.

Existing single-image 3DGS pipelines predict Gaussians from fixed image-grid locations, essentially one primitive per pixel. That works for renderings close to the original camera, but the authors argue those pixel-aligned primitives are only weakly coupled to actual scene surfaces, so they scatter and struggle to preserve coherent structures under large viewpoint shifts. InfiniSplat's contribution is to put the primitives on the surface itself: geometry-guided sampling places 2D supports according to depth-induced local surface structure, and a query-conditioned implicit decoder predicts Gaussian attributes from the image features queried at those supports.

Why bother if you are not a graphics researcher: single-image feed-forward 3D reconstruction is the piece that makes AR, robot scene understanding, and quick 3D asset generation economically tractable, since it avoids the cost of multi-view capture and per-scene optimization. If a cleaner geometric prior meaningfully improves how well the reconstruction holds up as you move the camera around, the technique matters for every downstream tool that has been getting by with pixel-aligned baselines.

The honest caveat is that the paper reports state-of-the-art performance on cross-dataset novel view synthesis evaluations and zero-shot generalization from Hypersim, a synthetic indoor dataset, to complex open-world scenes, but the abstract does not give you specific numeric deltas, runtime cost, or how the method behaves on reflective or dynamic content. Take the state-of-the-art claim as reported, not settled, until the community reproduces it against the released code.

The forward-looking part is that InfiniSplat is accepted to SIGGRAPH Asia 2026 (Journal Track) with code released, so labs working on neural rendering, and the AR/VR teams shopping for cheaper capture pipelines, get a concrete baseline to test surface-aligned Gaussians against their pixel-aligned incumbents.