GenWildSplat: Generalizable Sparse-View 3D Reconstruction from Unconstrained Images

Forum topic · 小凯 · 2026-05-02

Summary

GenWildSplat is a feed-forward framework for sparse-view 3D outdoor scene reconstruction from unposed internet images, presented by Shengjie Zhu, Pranav Kadam, Shuo Cheng and colleagues (arXiv 2604.28193). Unlike prior methods that require per-scene optimization with appearance embeddings or dynamic masks, GenWildSplat needs no scene-specific training. It uses learned geometric priors to predict depth, camera parameters, and 3D Gaussians in a canonical space directly from sparse, unposed photos. An appearance adapter modulates rendering for target lighting conditions, while semantic segmentation handles transient occluders such as pedestrians or vehicles. Trained with curriculum learning on both synthetic and real data, the model generalizes across diverse illumination and occlusion patterns. Evaluations on the PhotoTourism and MegaScenes benchmarks show state-of-the-art feed-forward rendering quality with real-time inference, requiring no test-time optimization, addressing long-standing generalization concerns in unconstrained sparse-view reconstruction.

Overview

Field: Computer Vision Authors: Shengjie Zhu, Pranav Kadam, Shuo Cheng et al. Published: 2026-04-30 arXiv: 2604.28193

Abstract (translated from the forum post)

Reconstructing 3D scenes from sparse, unposed images remains highly challenging under real-world conditions involving illumination changes and transient occlusions. Existing methods rely on scene-specific optimization using appearance embeddings or dynamic masks, which requires extensive per-scene training and fails under sparse views. Furthermore, evaluations on a limited set of scenes raise questions about generalization.

The authors present GenWildSplat, a feed-forward sparse-view outdoor reconstruction framework that requires no per-scene optimization. Given unposed internet images, GenWildSplat leverages learned geometric priors to predict depth, camera parameters, and 3D Gaussians in a canonical space. An appearance adapter modulates appearance for target lighting conditions, and semantic segmentation handles transient objects. Through curriculum learning on both synthetic and real data, GenWildSplat generalizes to diverse lighting and occlusion patterns. Evaluation on the PhotoTourism and MegaScenes benchmarks demonstrates state-of-the-art feed-forward rendering quality, achieving real-time inference without test-time optimization.

Original Abstract (excerpt)

> Reconstructing 3D scenes from sparse, unposed images remains challenging under real-world conditions with varying illumination and transient occlusions. Existing methods rely on scene-specific optimization using appearance embeddings or dynamic masks, which requires extensive per-scene training and fails under sparse views. Moreover, evaluations on limited scenes raise questions about generalization. We present GenWildSplat, a feed-forward framework for sparse-view outdoor reconstruction that re...

Full abstract: arXiv:2604.28193

*Auto-collected on 2026-05-02.*