NVIDIA Research Introduces ArtiFixer: Auto-Regressive Diffusion for Better 3D Gaussian Splatting

NVIDIA Research's Spatial Intelligence Lab has unveiled ArtiFixer, a two-stage pipeline that addresses a persistent weakness in 3D Gaussian Splatting (3DGS): poor extrapolation in under-observed areas. The method first fine-tunes a bidirectional video diffusion model using a novel opacity mixing strategy that encourages consistency with captured scene content while still generating plausible geometry in unseen regions. That bidirectional teacher is then distilled into a causal auto-regressive model capable of producing hundreds of frames in a single pass — eliminating the costly iterative distillation that limits existing approaches.

On standard benchmarks including MipNeRF 360 and DL3DV-10K, ArtiFixer reportedly outperforms all published baselines by 1–3 dB PSNR, handling scenarios where prior methods fail entirely. The work is accepted to SIGGRAPH 2026, with code and a Hugging Face demo already made available by the research team spanning NVIDIA, ETH Zurich, Cornell University, the University of Toronto, and the Vector Institute.