Difix3D+

Improving 3D Reconstructions with Single-Step Diffusion Models (CVPR 2025)

Difix3D+ - Improving 3D Reconstructions with Single-Step Diffusion Models (CVPR 2025)

Jay Zhangjie Wu, Yuxuan Zhang, Haithem Turki, Xuanchi Ren, Jun Gao, Mike Zheng Shou, Sanja Fidler, Zan Gojcic, Huan Ling

paper :
https://arxiv.org/abs/2503.01774
project website :
https://research.nvidia.com/labs/toronto-ai/difix3d/

핵심 :
nearly real-time single-step 2D diffusion model을
3D artifacts removing task에 맞게 fine-tune한 뒤,
3D model에 distill하여 progressively update하거나
post-processing으로 씀!

Contribution

Introduction

Overall Pipeline

Difix - from a Pretrained Diffusion Model to a 3D Artifact Fixer

Fine Tuning

위의 Overall Pipeline에서 Stage 1)에 해당되는 내용!

noise level이 높으면 model은 artifacts를 잘 제거하지만 image context도 함께 바꿈,, noise level이 낮으면 model은 image를 거의 안 건드림

Data Curation

Difix3D+ - NVS with Diffusion Priors

Difix3D - Progressive 3D Update

위의 Overall Pipeline에서 Stage 2)의 Step 1), 2)에 해당되는 내용!

Difix3D+ - Real time Post Render Processing

위의 Overall Pipeline에서 Stage 2)의 Step 3)에 해당되는 내용!

Experiment

In-the-wild Artifact Removal

Automotive Scene Enhancement

Ablation Study

Conclusion

Question