Jonathan Lorraine#1108
The ArtiFixer code and model weights are now released! Links to both on our project page: https://research.nvidia.com/labs/sil/projects/artifixer/
Per-scene methods like 3DGS look great on captured views but collapse off-trajectory. We repair those artifacts and beat prior SOTA by 1–3 dB PSNR. 🧵(1/6)
Jonathan Lorraine#1108The ArtiFixer code and model weights are now released! Links to both on our project page: https://research.nvidia.com/labs/sil/projects/artifixer/
Per-scene methods like 3DGS look great on captured views but collapse off-trajectory. We repair those artifacts and beat prior SOTA by 1–3 dB PSNR. 🧵(1/6)
Users praised NVIDIA's release of ArtiFixer code and weights to fix 3DGS artifacts because it addresses issues they've struggled with and seems effective enough to try immediately.
(2/6) Two stages:
1. A bidirectional teacher video model, finetuned with opacity mixing — blend rendered-RGB latents with noise by the rendered opacity. Faithful where observed, generative where not.
2. Distill it into a causal AR model that generates 100s of frames per pass.
(6/6) If you're at CVPR - be sure to catch @RdeLutio's talk at the GenRecon3D workshop today! https://genrecon3d.github.io/
Many thanks to all of our amazing co-authors! @TobiasFischer11 Yen-Yu Chang Yuxuan Zhang @jayzhangjiewu @xuanchi13 @TianchangS @KatTothova @ZGojcic
(4/6) On sparse DL3DV:
- GenFusion (https://genfusion.sibowu.com/) generates 16 frames at a time → iterative distillation → blur. - Gen3C (https://research.nvidia.com/labs/toronto-ai/GEN3C/) is sharp but ignores the source content. - Our method is sharp AND consistent, even when the input is degraded.