ArtiFixer

ArtiFixer is a research project from NVIDIA that improves 3D scene reconstructions by filling in visual errors and generating views from camera angles that were never photographed. When you take a set of photos of a real place and run them through a 3D reconstruction pipeline, the resulting model often has blurry patches, missing geometry, or visual artifacts in areas where photo coverage was thin. ArtiFixer takes a base 3D reconstruction and uses an AI model to correct those problems.

The underlying approach uses a type of AI model called a diffusion model (the same family of models used for AI image generation) paired with a technique that generates frames one at a time in sequence, so that each new frame is informed by what came before. This allows the model to produce consistent, artifact-free views even along camera paths that were never in the original photo set. The result is a corrected or extended 3D reconstruction that can be viewed from novel angles.

Using the project requires a GPU with CUDA support, Docker, and a working 3D reconstruction from a tool called COLMAP, which estimates camera positions from a set of photos. You provide your photos, run COLMAP to get a sparse reconstruction, then use the provided preparation scripts to format the data for ArtiFixer. After that, a single inference command runs the model and saves corrected frames. A pre-trained 14-billion-parameter model checkpoint is available on Hugging Face (NVIDIA's model hosting page for it).

This is the official code release for a paper presented at SIGGRAPH 2026, the main computer graphics research conference. It is published by NVIDIA's research labs and licensed under Apache 2.0. The repository includes code for model training, inference, data preparation, and evaluation on standard 3D scene datasets.