nv-tlabs/cosmos-transfer1-diffusion-renderer

Decouples inverse rendering (albedo, normal, depth, roughness, metallic estimation) from forward rendering using separate diffusion transformers, enabling flexible relighting with custom environment maps or randomized illumination. Built on NVIDIA's Cosmos World Foundation Models with nvdiffrast for differentiable rendering, supporting both image and video inputs with configurable G-buffer passes. Targets synthetic data generation for training robust vision and robotic systems under varying lighting conditions.

786 stars. No commits in the last 6 months.