BRDFusion: Combining PBR and Generative Models

The new BRDFusion framework enables the reconstruction of geometry and lighting in urban scenes from video using 3D Gaussian Splatting and diffusion models.

Compiled by Sergey KostenchukPublished 2026-06-22Updated 2026-06-22

2026-06-22 Multimodal HuggingFace

🎨 BRDFusion: Where Physics Meets Generation

A new framework combines physically based rendering (PBR) and generative models for the inverse rendering of urban scenes. The system reconstructs geometry, materials, and HDR lighting directly from video, utilizing 3DGS and diffusion models.

🌍 This technology enables the creation of high-quality simulations for autonomous driving and content creators, bridging the gap between rendering accuracy and generative realism.

👤 Users will be able to edit urban scene videos by changing the time of day or adding objects while maintaining physical correctness.

Source 1: https://shigon255.github.io/brdfusion-page/ Source 2: https://github.com/shigon255/BRDFusion

Sources