# TripoSplat
TripoSplat is an image-to-3D model by TripoAI (VAST). It takes one picture of an object and returns the whole object as 3D Gaussian splats. It is the model Speridlabs Idol builds on, and LibreYOLO runs it through the same LibreGen call and output.
Tasks: 3D generation. Install: pip install "libreyolo[gen]".
Verified against LibreYOLO v1.8.0.

## Install

```bash
pip install "libreyolo[gen]"
```

The extra brings `diffusers` (FLUX.2 VAE) and `transformers` (DINOv3 encoder).

## Predict

Nothing is gated and no login is needed. Weights download on first use at
pinned revisions.

**Python**

```python
from libreyolo import LibreGen, SAMPLE_IMAGE

model = LibreGen("triposplat")      # weights download on first use
result = model(SAMPLE_IMAGE, seed=0)

result.save("object.ply")           # standard 3DGS layout
result.save("object.splat")         # compact web format
print(len(result.gaussians))
```

**CLI**

```bash
# Saves the object as .ply under runs/generate/predict.
libreyolo generate source=https://raw.githubusercontent.com/LibreYOLO/libreyolo/release/libreyolo/assets/guggenheim-bilbao.jpg model=triposplat seed=0
```

**Settings**

```python
from libreyolo import LibreGen, SAMPLE_IMAGE

model = LibreGen("triposplat")
result = model(SAMPLE_IMAGE, steps=20, gaussians=131072)
```

A 0.66B rectified-flow transformer denoises an object latent and the input
camera, conditioned on DINOv3 and FLUX.2 VAE tokens of the 1024 px input. An
octree decoder places anchors and a Gaussian decoder turns each anchor into 32
Gaussians, from 32,768 to 262,144 in all. It runs in fp16 on CUDA and Apple
MPS and in fp32 on CPU, with no compiled path.

The output is the same as [Speridlabs Idol](/docs/models/speridlabs-idol):
`result.gaussians` with `xyz`, `scales`, `rotations`, `opacities`, `sh`,
`colors`, the inferred `camera`, `seed` and `params`. The object is z-up and
centred in `[-0.5, 0.5]^3`. `.ply` is the standard 3DGS layout turned y-up;
`.splat` is the compact web format.

The seed drives a generator on the model's device, as upstream does, so one
seed gives different objects on different devices. Unlike upstream, the octree
sampling also draws from the seeded generator, so a seed reproduces the whole
object.

### Settings

| Setting | Default | Meaning |
| --- | --- | --- |
| `steps` | `20` | Sampling steps; upstream recommends 10 to 20. |
| `guidance` | `3.0` | Classifier-free guidance; 1 or less turns it off. |
| `shift` | `3.0` | Flow time shift. |
| `gaussians` | `262144` | Gaussians in the output, from 32,768 to 262,144. |
| `margin` | `1.2` | Crop side over object size, as upstream. |

## Variants

One size, `0.6b` (0.66B parameters), about 1.8 GB plus the DINOv3 ViT-H+
encoder (1.7 GB) and the FLUX.2 VAE encoder (0.2 GB) that Idol also uses.

## Export

| Task | ONNX | TorchScript | ExecuTorch | TensorRT | OpenVINO | Paddle | MNN | RKNN | Axelera | ncnn | TFLite | CoreML | Core AI |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 3D generation |  |  |  |  |  |  |  |  |  |  |  |  |  |

A "yes" means the export is supported. An empty cell means the exporter refuses that combination.

Export is not implemented for LibreGen 3D families. `export()`, `train()` and
`val()` raise `NotImplementedError` (ADR 0031 in the library).

## Licensing

Check the license on the Hugging Face repository of the specific weights you download. That repository is authoritative and licenses are not always uniform across a family. This is a description of the licenses involved, not legal advice.

- Original work: TripoSplat, VAST (TripoAI)
- Upstream license: MIT
- Upstream source: https://github.com/VAST-AI-Research/TripoSplat
- LibreYOLO code: MIT and Apache-2.0
- Weights: MIT, republished at https://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 denoiser, octree and Gaussian decoders it ports from TripoSplat keep MIT, and the preprocessing and encoder tokenization shared with Speridlabs Idol come from the Apache-2.0 Idol SDK, with notices in libreyolo/models/gen/triposplat/NOTICE. The denoiser, decoder and BiRefNet weights are redistributed by LibreYOLO as LibreYOLO/LibreTripoSplat0.6b (MIT), byte-identical to VAST-AI/TripoSplat, and the FLUX.2 VAE downloads from black-forest-labs/FLUX.2-klein-4B (Apache-2.0). The DINOv3 image encoder is the exception: it is Meta's, under the DINOv3 License, a custom agreement that is not OSI-approved. It permits redistribution with a copy of the agreement and passes through trade-control and end-use restrictions (military, weapons, espionage and nuclear uses among them). LibreYOLO mirrors that encoder in bfloat16 at LibreYOLO/LibreDINOv3vith16plus together with the agreement, so nothing is gated; read the agreement before shipping a product on it.

TripoSplat is by TripoAI
([VAST-AI-Research/TripoSplat](https://github.com/VAST-AI-Research/TripoSplat)).
Code and weights are MIT; LibreYOLO mirrors the denoiser, decoder and BiRefNet
files byte-identical as LibreYOLO/LibreTripoSplat0.6b. The FLUX.2 VAE comes
from black-forest-labs/FLUX.2-klein-4B under Apache-2.0.

**The DINOv3 image encoder is not MIT.** It is Meta's, under the DINOv3
License, a custom agreement with trade-control and end-use restrictions
(military, weapons, espionage and nuclear uses among them). LibreYOLO mirrors
it in bfloat16 at LibreYOLO/LibreDINOv3vith16plus with the agreement, so
nothing is gated. Read it before shipping a product on this model.

## Citation

```bibtex
@misc{yan2026generative3dgaussianslearned,
    title={Generative 3D Gaussians with Learned Density Control}, 
    author={Runjie Yan and Yan-Pei Cao and Peng Wang and Ding Liang and Yuan-Chen Guo},
    year={2026},
    eprint={2605.16355},
    archivePrefix={arXiv},
    primaryClass={cs.GR},
    url={https://arxiv.org/abs/2605.16355}, 
}
```

Copied from https://github.com/VAST-AI-Research/TripoSplat#citation
