LibreTripoSR
TripoSR, the fast feed-forward image-to-3D model by Tripo AI and Stability AI ("TripoSR: Fast 3D Object Reconstruction from a Single Image", 2024), repackaged for LibreYOLO. One picture of an object in, a vertex-coloured triangle mesh out.
LibreYOLO runs it in plain PyTorch and meshes it with scikit-image, so the same model runs on NVIDIA GPUs, Apple silicon (MPS) and CPU.
from libreyolo import LibreGen
model = LibreGen("triposr") # downloads this repo
model("chair.png").save("chair.obj")
libreyolo generate source=chair.png model=triposr
This is a mirror, not a new model
The TripoSR weights are Tripo AI's and Stability AI's. Nothing was retrained,
converted or re-encoded; config.yaml and model.ckpt are byte-identical to
upstream.
- Upstream repository:
stabilityai/TripoSR - Upstream revision:
5b521936b01fbe1890f6f9baed0254ab6351c04a - Code: VAST-AI-Research/TripoSR
- Paper: arXiv:2403.02151
| File | Contents |
|---|---|
config.yaml |
upstream's model configuration |
model.ckpt |
the whole network: DINO ViT-B/16 image encoder, triplane transformer, upsampler, NeRF MLP |
Upstream's demo videos are not mirrored. LibreYOLO removes backgrounds with
BiRefNet (MIT), loaded from
LibreYOLO/LibreTRELLISl.
SHA-256 of every file is listed in NOTICE.
Training data and use
Upstream trained TripoSR on renders of a curated subset of Objaverse, available under the CC-BY license. Upstream's card asks that the model not be used to intentionally create or disseminate 3D models that people would foreseeably find disturbing, distressing, or offensive, or content that propagates historical or current stereotypes.
License
MIT, as published by Tripo AI and Stability AI. See LICENSE and NOTICE.