TripoSR

TripoSR generates a textured 3D mesh from a single photograph in under half a second on an NVIDIA A100, a joint release from Tripo AI and Stability AI. The model follows the Large Reconstruction Model approach, using a feed-forward transformer rather than iterative optimization, which is what makes inference so much faster than score-distillation methods. Output can be exported with vertex colors or full textures, and the repository includes batch processing, a Gradio web interface, and CUDA-accelerated marching cubes via torchmcubes for mesh extraction. Running it takes a GPU with roughly 6 GB of VRAM, PyTorch, and Python 3.8 or newer. Both the code and pretrained weights are released under the MIT license, so commercial use is permitted. Game developers, 3D artists, and researchers use it to prototype assets and study single-view reconstruction, and the authors report it outperforming other open source image-to-3D models on public benchmarks.