Reproduce “BRDFusion: Physics Meets Generation for Urban Scene Inverse Rendering” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yi-Ruei Liu, Jie-Ying Lee, Zheng-Hui Huang, Yu-Lun Liu, Chih-Hao Lin
BRDFusion integrates physically-based rendering with generative diffusion models to perform inverse rendering on dynamic urban scenes, recovering 3D geometry, materials, and lighting from video. This hybrid approach yields photorealistic and controllable scene reconstructions, demonstrating superior material property estimation and significantly improved relighting accuracy compared to existing methods.
Key insights
- Careful integration of generative priors from diffusion models is essential for regularizing the ill-posed inverse rendering problem, preventing artifacts and improving the accuracy of geometry, material, and lighting decomposition.
- A multi-stage optimization pipeline, including generative refinement of intrinsic G-buffer maps, significantly enhances the detail, physical plausibility, and temporal consistency of recovered 3D scene properties.
- Combining a physically-based rendering pass with a generative denoising pass within the forward rendering pipeline ensures both precise light transport control and high photorealistic quality in novel view synthesis and relighting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.