nvdiffrec

Why we included this project

This is the official implementation of the CVPR 2022 paper that recovers a complete triangular mesh, its materials, and an environment light from ordinary multi-view photographs. Rather than stopping at a point cloud or implicit field, it optimizes topology, texture, and lighting together, so the result is a clean asset that can go straight into a game or renderer. It relies on differentiable rendering with custom CUDA extensions, and the slang branch replaces the hand-written gradients with automatic differentiation via slangpy while keeping the same runtime performance. Anyone working on photogrammetry or neural reconstruction will find it a useful reference for how geometry and appearance can be solved jointly from plain 2D supervision. FlexiCubes isosurfacing is also supported if you want to try a different topology representation.