70x speedup while maintaining image quality, KV caching + fewer-step sampling directly allows a single card to run a 14B model, this distillation approach is too aggressive.

NVIDIA and ETH Zurich launched the ArtiFixer3D reconstruction framework at SIGGRAPH 2026.
By using distribution matching distillation to convert bidirectional video large models into autoregressive models, it improves reconstruction quality and speeds up by about 70 times, with image quality surpassing state-of-the-art benchmarks by 1–3 dB, and can generate hundreds of frames in a single inference using KV caching and few-step sampling.
In single-GPU GB300 testing, the 14B version achieves 8.36 fps, and the 1.3B version reaches 34.38 fps.
Three variants are provided to optimize sharpness and consistency; in the DL3DV blind spot test, ArtiFixer3D+ PSNR is 20.15 dB, approximately 3.12 dB ahead of GenFusion.