InfiniSplat Moves Single-Image 3D Gaussian Splatting Toward Surface-Aligned Geometry

Introduction

Turning a single photograph into a 3D scene that can be rendered from new viewpoints is one of the central ambitions of modern view synthesis. 3D Gaussian Splatting has become a popular representation because it can deliver efficient rendering and visually strong results. Yet many successful 3DGS pipelines still depend on multi-view capture or per-scene optimization, both of which limit accessibility and scalability.

InfiniSplat addresses the more constrained setting: feed-forward 3D Gaussian Splatting from just one image. The work, accepted to ACM Transactions on Graphics as part of the SIGGRAPH Asia 2026 Journal Track, focuses on a key weakness of existing single-image methods: they often look plausible for nearby views but become structurally unstable when the viewpoint changes significantly.

Key ideas

• The limitation of pixel-aligned primitives
  Many feed-forward single-image 3DGS approaches predict Gaussians from fixed image-grid locations. This pixel-aligned design is convenient, but it does not necessarily match the geometry of the underlying scene. As a result, the generated primitives may be only weakly coupled to real surfaces and can scatter or lose coherence under large-baseline camera motion.

• A shift toward surface-aligned representation
  InfiniSplat reframes the placement of Gaussian supports. Rather than treating pixel centers as the default anchors, it uses geometry-guided sampling based on depth-induced local surface structure. The goal is to make support locations better reflect the shape and orientation of visible scene surfaces.

• Implicit Gaussian decoding
  After support points are selected, InfiniSplat queries image features at these locations and feeds them into a query-conditioned implicit decoder. This decoder predicts the Gaussian attributes needed for rendering. By decoupling attribute prediction from fixed pixel centers, the framework can produce Gaussian layouts that more closely follow scene geometry.

• Large-baseline novel view synthesis
  The reported evaluations focus on cross-dataset novel view synthesis. According to the paper summary, InfiniSplat achieves state-of-the-art performance compared with single-image feed-forward baselines and demonstrates zero-shot generalization from Hypersim indoor synthetic training to complex open-world scenes.

Why it matters

The main contribution is representational. InfiniSplat suggests that the bottleneck in single-image 3DGS is not only the amount of input information, but also how predicted primitives are anchored. If Gaussian primitives remain locked to the 2D image grid, the model can struggle to form a coherent 3D structure when the viewer moves far from the original camera.

A surface-aligned strategy could make single-image 3D reconstruction more useful for AR/VR, spatial content generation, photo-to-3D tools, and rapid scene visualization. The feed-forward design avoids per-scene optimization, while the single-image setup reduces capture requirements. The paper summary does not provide detailed numerical metrics, so the strongest claims should still be checked against the full paper, code, and demo.

Source: Hugging Face Daily Papers