Researchers have introduced InfiniSplat, a new framework for feed-forward single-image 3D Gaussian Splatting (3DGS) that shifts from pixel-aligned to surface-aligned representations. The work, accepted to ACM Transactions on Graphics (SIGGRAPH Asia 2026 Journal Track), addresses a key limitation of existing methods: their reliance on fixed image-grid locations for Gaussian prediction, which often fails to preserve coherent structures under large viewpoint shifts.
InfiniSplat constructs its representation in two stages. First, it uses geometry-guided sampling to place 2D supports according to depth-induced local surface structure. Then, a query-conditioned implicit decoder predicts Gaussian attributes from image features queried at these supports. By grounding support locations in geometry and decoupling Gaussian prediction from fixed pixel centers, the method produces Gaussian layouts that better follow scene surfaces and reduce scattered primitives caused by grid discretization.
In cross-dataset novel view synthesis evaluations, InfiniSplat achieves state-of-the-art performance compared with other single-image feed-forward baselines. It also demonstrates zero-shot generalization from Hypersim indoor synthetic training to complex open-world scenes.
The project page, code, and a live demo are available online, providing resources for researchers and practitioners to explore the method.