When 3D models see things that aren't there

3D foundation models like COLMAP and MASt3R can fool standard evaluation metrics by hallucinating consistent geometry from images that don't depict the same scene—noise, repeated frames, or completely unrelated views. Researchers benchmarked this failure mode and decomposed existing neural metrics to understand why they break. They built geometric verification metrics using classical computer vision that correlate 4× better with human judgment on real 3D reconstruction outputs.