TripoSR
TripoSR is a feed-forward image-to-3D model by Tripo AI and Stability AI. It turns one picture of an object into a triplane NeRF in a single pass and returns the object as a vertex-coloured triangle mesh. LibreYOLO runs it through the LibreGen factory in plain PyTorch.
- Tasks
- 3d generation
- Install
pip install "libreyolo[gen]"- Support tier
- Sibling tier, since v1.8.0. A separate product surface with its own factory and contract.
- Licenses
- Code MIT and Apache-2.0, weights MIT. Commercial use
Install
TripoSR runs on the LibreGen extra, which brings scikit-image for marching cubes.
pip install "libreyolo[gen]"Predict
Weights download on first use from LibreYOLO/LibreTripoSR, with nothing gated and no login.
from libreyolo import LibreGen, SAMPLE_IMAGE model = LibreGen("triposr") # weights download on first useresult = model(SAMPLE_IMAGE) result.save("object.obj") # vertex colours as `v x y z r g b`m = result.meshesprint(m.vertices[0].shape, m.faces[0].shape) # (V, 3), (F, 3)# Saves the object as .obj under runs/generate/predict.libreyolo generate source=https://raw.githubusercontent.com/LibreYOLO/libreyolo/release/libreyolo/assets/guggenheim-bilbao.jpg model=triposrfrom libreyolo import LibreGen, SAMPLE_IMAGE model = LibreGen("triposr")result = model(SAMPLE_IMAGE, resolution=384) # finer marching-cubes gridresult = model(SAMPLE_IMAGE, threshold=25.0) # density at the surfaceA DINO ViT-B/16 encodes the 512 px input, a transformer decodes a triplane
NeRF from it in one pass, and the mesh is the marching-cubes isosurface of the
NeRF's density. There is no sampling, so seed is accepted like every
LibreGen family but changes nothing. It takes about a second on a GPU and runs
in fp32 on NVIDIA GPUs, Apple silicon and CPU.
The source is anything predict takes: a path, URL, PIL image, NumPy array, a folder or a list. An image with real transparency is used as the cutout; anything else has its background removed first with LibreYOLO's BiRefNet port.
result.meshes holds one mesh per object: vertices (V, 3), z-up inside
[-0.5, 0.5]^3, faces (F, 3), and extras["vertex_colors"] (V, 3) RGB in
[0, 1]. The input view looks along -x, so the front of the object faces +x.
save() writes .obj with vertex colours.
Settings
| Setting | Default | Meaning |
|---|---|---|
resolution | 256 | Marching-cubes grid, from 32 to 1024. |
threshold | 25.0 | Density at the surface. |
margin | 1/0.85 | Crop side over object size (upstream's foreground ratio 0.85). |
Marching cubes runs through scikit-image instead of upstream's torchmcubes: the same isosurface, triangulated differently.
Variants
One checkpoint, 0.42B parameters, about 1.7 GB plus 0.4 GB for BiRefNet.
Export
| Task | ONNX | TorchScript | ExecuTorch | TensorRT | OpenVINO | Paddle | MNN | RKNN | Axelera | ncnn | TFLite | CoreML | Core AI |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 3D generation | 3D generation to ONNX: not supported | 3D generation to TorchScript: not supported | 3D generation to ExecuTorch: not supported | 3D generation to TensorRT: not supported | 3D generation to OpenVINO: not supported | 3D generation to Paddle: not supported | 3D generation to MNN: not supported | 3D generation to RKNN: not supported | 3D generation to Axelera: not supported | 3D generation to ncnn: not supported | 3D generation to TFLite: not supported | 3D generation to CoreML: not supported | 3D generation to Core AI: not supported |
Export is not implemented for LibreGen 3D families. export(), train() and
val() raise NotImplementedError; their names are reserved by ADR 0031 in
the library.
Licensing
Check the license on the Hugging Face repository of the specific weights you download. Every checkpoint in the LibreYOLO org carries one, and they are not always the same across a family. That repository is the authoritative source; the summary below describes what applied when this page was last verified.
This is a description of the licenses involved, not legal advice. If the answer matters commercially, read the licenses yourself and take your own counsel.
- Original work
- TripoSR, Tripo AI and Stability AI
- Upstream license
- MIT
- Upstream source
- github.com/VAST-AI-Research/TripoSR
- LibreYOLO code
- MIT
- Weights
- MIT, republished at huggingface.co/LibreYOLO
- Interpretation
- MIT is a permissive license, so the code and weights can be used in commercial and closed-source products; the only obligation is to keep the copyright and license notice with copies. LibreYOLO's adapter is MIT; the triplane, NeRF and mesh-extraction code it adapts from TripoSR keeps MIT, and the transformer and DINO ViT encoder follow HuggingFace diffusers and transformers code under Apache-2.0, with their notices in libreyolo/models/gen/triposr/NOTICE. LibreYOLO avoids upstream's torchmcubes (MPL-2.0) by running marching cubes through scikit-image (BSD-3-Clause). The stabilityai/TripoSR weights (MIT) are redistributed unchanged as LibreYOLO/LibreTripoSR; Background removal runs on LibreYOLO's BiRefNet port with the ZhengPeng7/BiRefNet weights (MIT), from LibreYOLO/LibreTRELLISl. Both are pinned and not gated.
TripoSR is by Tripo AI and Stability AI. The code
(VAST-AI-Research/TripoSR) and
the stabilityai/TripoSR weights are MIT, so commercial use is fine with the
copyright notice kept. LibreYOLO redistributes the weights unchanged as
LibreYOLO/LibreTripoSR. It avoids upstream's torchmcubes (MPL-2.0) by using
scikit-image (BSD-3-Clause); the BiRefNet background remover weights are MIT.
Citation
@article{TripoSR2024,
title={TripoSR: Fast 3D Object Reconstruction from a Single Image},
author={Tochilkin, Dmitry and Pankratz, David and Liu, Zexiang and Huang, Zixuan and and Letts, Adam and Li, Yangguang and Liang, Ding and Laforte, Christian and Jampani, Varun and Cao, Yan-Pei},
journal={arXiv preprint arXiv:2403.02151},
year={2024}
}Copied from the authors' citation block at github.com/VAST-AI-Research/TripoSR#citation.