TripoSR is a useful starting point for turning a single object image into a mesh you can inspect and clean up locally. Its 2024 release from Tripo and Stability AI includes MIT-licensed code and pretrained models, with vertex-color output and optional texture baking. Choose it for prop prototypes or for learning how reconstruction fits into an import-and-cleanup workflow. Unseen surfaces are inferred, and a finished character still needs conventional modeling and rigging work. The authors' generation-speed claim uses an NVIDIA A100, so expect your hardware to set the pace. This entry covers TripoSR's local release; Tripo's hosted service is a separate product.
Commercial use & license
MIT covers the cited code and pretrained release. Retain required notices, and check preprocessing dependencies and rights to input images separately.
Adult content rules
The cited MIT license contains no adult-content restriction, but it does not grant rights to reference images, people's likenesses or third-party assets. Check the exact checkpoints and dependencies; hosted services have separate terms. This is not blanket permission for adult use.
What you need
Python 3.8+, PyTorch and the repository's dependencies. Default single-image settings are documented at about 6 GB VRAM; the script falls back to CPU when CUDA is unavailable. The 6 GB figure is a reported configuration, not a verified minimum.
Ranks are recommendations for the task described, based on reviewed documentation. Model ranks cover non-explicit development work. We have not run practical benchmarks. Check content policies and commercial terms separately. How the ranks work →
Sources & further reading +
Last checked 5 October 2026.