ArtiFixer: Enhancing and Extending 3D Reconstruction with Auto-Regressive Diffusion Models

Researchers propose ArtiFixer, a two-stage pipeline using auto-regressive diffusion models to enhance 3D reconstruction quality. The method addresses scalability and quality issues in existing approaches by training a bidirectional generative model with opacity mixing, then distilling it into a causal auto-regressive model that generates hundreds of frames in a single pass.

• →ArtiFixer solves scalability issues in 3D reconstruction by generating hundreds of views in a single pass rather than requiring costly iterative processes.
• →The method uses a novel opacity mixing strategy to maintain consistency with existing observations while extrapolating content in unseen areas.
• →A two-stage pipeline first trains a bidirectional generative model, then distills it into an efficient auto-regressive model.
• →The approach outperforms existing baselines by 1-3 dB PSNR on commonly benchmarked datasets.
• →The method can generate plausible reconstructions in scenarios where current approaches fail completely.