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

2026-07-06

AI Summary

Existing feedforward 3D reconstruction methods struggle to achieve high-quality novel view synthesis in uncontrolled outdoor scenes with drastic lighting variations. This work proposes WildSplat, the first feedforward Gaussian splatting framework tailored for pose-free wild images. WildSplat introduces a geometry-appearance disentangled dual-branch architecture, a global pre-modulated cross-attention mechanism, and a multi-reference joint training strategy to explicitly model geometry and appearance while enabling conditional rendering. For the first time, it achieves appearance-controllable, single-pass novel view synthesis from pose-free wild images. Under sparse input settings, WildSplat significantly outperforms both optimization-based and feedforward baselines, establishing a new state of the art in this challenging domain.

Abstract

While feedforward 3D reconstruction excels at efficient novel view synthesis, it typically falters when faced with scenes under varying illumination. To this end, we introduce WildSplat, the first feedforward 3D Gaussian Splatting framework capable of appearance-conditioned novel-view synthesis for unposed in-the-wild images. To handle inconsistent photometric conditions, we propose a dual-branch architecture that explicitly decouples geometry from appearance. The geometry branch extracts an appearance-invariant 3D structure and jointly predicts camera poses. To govern the rendering appearance, the appearance branch injects target appearance cues into the content features via a globally pre-modulated cross-attention mechanism. To further prevent feature entanglement, we introduce a joint multi-reference training strategy that stabilizes the training process. Extensive experiments show that WildSplat surpasses existing optimization-based and feedforward methods, achieving state-of-the-art performance in in-the-wild novel view synthesis and appearance editing from sparse inputs in a single forward pass.

Problem

• feedforward 3D reconstruction
• in-the-wild images
• novel view synthesis
• appearance conditioning
• varying illumination

Innovation

• Gaussian Splatting
• feedforward 3D reconstruction
• appearance-conditioned synthesis
• geometry-appearance disentanglement
• in-the-wild images