WildSplatter: Feed-forward 3D Gaussian Splatting with Appearance Control from Unconstrained Images

2026-04-24 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

WildSplatter is a novel feed-forward 3D Gaussian Splatting (3DGS) model designed to process unconstrained images with unknown camera parameters and varying lighting conditions. Unlike traditional 3DGS, which requires iterative optimization and consistent lighting, WildSplatter is trained on diverse photo collections like MegaScenes. It jointly learns 3D Gaussians and appearance embeddings, allowing flexible modulation of Gaussian colors to represent significant variations in lighting and appearance. The model reconstructs 3D Gaussians from sparse input views in under one second, enabling real-time novel view synthesis and appearance control. Experimental results on the NeRF-OSR dataset demonstrate that WildSplatter outperforms existing pose-free 3DGS methods, achieving higher PSNR and lower LPIPS scores, particularly when handling diverse illumination.

Key takeaway

For research scientists developing real-time 3D scene reconstruction and novel view synthesis systems, WildSplatter offers a significant advancement. You should consider integrating its feed-forward, pose-free approach to handle unconstrained image collections with varying lighting, as it provides superior performance and appearance control compared to existing methods, without the computational overhead of iterative optimization.

Key insights

WildSplatter enables fast, pose-free 3DGS from unconstrained images by disentangling geometry and appearance via learned embeddings.

Principles

• Geometry remains invariant under lighting changes.
• Appearance variations can be captured in a low-dimensional latent space.

Method

WildSplatter uses a Vision Transformer backbone and dual-DPT heads to estimate colorless 3D Gaussian geometry and local scene features. Global appearance embeddings are then predicted from target images and combined with local features to determine Gaussian colors.

In practice

• Reconstruct 3D Gaussians from sparse views in under one second.
• Control appearance via embedding interpolation.
• Generalize appearance control across different datasets.

Topics

• 3D Gaussian Splatting
• Feed-forward Models
• Appearance Embeddings
• Pose-free Reconstruction
• Unconstrained Images

Code references

• yfujimura/WildSplatter

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• GlobalSplat: Efficient Feed-Forward 3D Gaussian Splatting via Global Scene Tokens — GlobalSplat uses a global latent scene representation for efficient, compact, and consistent 3D Gaussian Splatting.
• Feed-Forward 3D Scene Modeling: A Problem-Driven Perspective — Feed-forward 3D reconstruction offers efficient, generalizable scene modeling by directly mapping 2D inputs to 3D representations.
• WildDet3D - an open model for monocular 3D detection — WildDet3D predicts 3D object bounding boxes from single RGB images across 13K+ categories without fine-tuning.
• UFO-4D: Unposed Feedforward 4D Reconstruction from Two Images — UFO-4D reconstructs dense 4D representations from two unposed images using dynamic 3D Gaussian Splats.
• Less Gaussians, Texture More: 4K Feed-Forward Textured Splatting — LGTM enables 4K novel view synthesis by decoupling geometry from resolution using textured Gaussian primitives.
• EnerGS: Energy-Based Gaussian Splatting with Partial Geometric Priors — EnerGS uses a soft energy field from partial geometry to guide 3D Gaussian Splatting, improving reconstruction quality.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.