GenWildSplat Reconstructs 3D Scenes from Sparse, Unconstrained Images in Under 3 Seconds

Researchers from UIUC, UMD, and collaborating institutions have introduced GenWildSplat, a feed-forward method for reconstructing 3D scenes from as few as two unconstrained, unposed images using 3D Gaussian Splatting. Accepted to CVPR 2026, the system completes reconstruction in roughly three seconds on a single A6000 GPU, handling real-world challenges such as varying illumination, transient objects, and sparse viewpoints without requiring per-scene optimization.

A key differentiator is the model's decoupled handling of appearance and geometry: a dedicated light encoder captures per-image illumination, enabling flexible novel-view synthesis under different lighting conditions and even cross-scene appearance transfer — capabilities that prior methods like WildGaussians and NexusSplats do not support. Reconstruction quality scales with the number of input views (tested up to six), and the framework generalizes to scenes not seen during training, as demonstrated on the MegaScenes outdoor dataset.