What is ambisur?

AmbiSuR is a Python research codebase presented at ICML 2026 that tackles a tricky problem in 3D surface reconstruction from photographs: photometric ambiguity. Gaussian Splatting represents a 3D scene as a collection of tiny translucent "splat" shapes (called Gaussians) that together describe how the scene looks from any camera angle. In challenging capture conditions, appearance variations across photos make it hard for the algorithm to pinpoint where real surfaces are, producing distorted or incorrect geometry. AmbiSuR introduces two main mechanisms to detect and correct for this ambiguity. Gaussian Primitive Truncation removes shape primitives that are likely incorrectly placed due to photometric inconsistency. Ray-Color Consistency regularization encourages color agreement across different camera viewpoints. The system also integrates depth priors from Depth Anything 3, which provides initial distance estimates anchored to COLMAP-format camera data, stabilizing geometry before full optimization. Getting the code running starts with creating a conda environment from a provided environment.yml file, and requires PyTorch 2.0 or newer for compatibility with Depth Anything 3. A separate set of PyTorch-related and custom CUDA components is installed afterward, and the Depth Anything 3 package itself is an optional install for those who want the multi-view depth priors. The workflow is: prepare photos with COLMAP to get camera positions, run a depth prior script, train the Gaussian scene representation, then extract a 3D mesh. Configurable hyperparameters control how aggressively ambiguous Gaussians are pruned and how much the depth prior influences results. Evaluation scripts are provided for three standard benchmarks: DTU, Tanks and Temples, and Mip-NeRF 360. The project also has a dedicated project page and an arXiv paper alongside the code. This targets researchers and practitioners working on 3D computer vision. The full README is longer than what was provided.