NVlabs/nvdiffrec

Performs joint differentiable optimization of mesh topology, physically-based materials (albedo, roughness, metallic), and environment lighting from multi-view images using differentiable marching tetrahedra for geometry and custom CUDA rendering kernels. Built on PyTorch with tight integration to NVIDIA's rendering stack (nvdiffrast, tiny-cuda-nn), supporting both single-GPU and multi-GPU DDP training on high-end GPUs; also offers a Slang-based autodiff variant for simplified shader-based rendering.

2,270 stars. No commits in the last 6 months.