WildSplat: Feedforward Gaussian Splatting from Unposed In-the-Wild Images

Abstract

Feedforward 3D reconstruction enables efficient novel-view synthesis, but it typically assumes photometric consistency across input views. In-the-wild photo collections violate this assumption: images of the same scene may vary dramatically in illumination, weather, exposure, and capture time.

WildSplat is a feedforward 3D Gaussian Splatting framework for appearance-conditioned novel-view synthesis from sparse, unposed, in-the-wild images. It uses a dual-branch architecture to decouple appearance-invariant geometry from target-conditioned appearance: the geometry branch predicts camera poses and 3D Gaussian structure, while the appearance branch injects reference appearance cues through globally pre-modulated cross-attention. A joint multi-reference training strategy further discourages geometry-appearance entanglement, enabling state-of-the-art in-the-wild synthesis and appearance editing in a single forward pass.

BibTeX

@inproceedings{zhang2026wildsplat,
  title={WildSplat: Feedforward Gaussian Splatting from Unposed In-the-Wild Images},
  author={Zhang, Xiyu and Zhuang, Jingyu and Zhai, Hongjia and Yan, Zizheng and Chen, Jinwei and Zhang, Guofeng and Fan, Qingnan},
  booktitle={European Conference on Computer Vision},
  year={2026}
}