LibreTripoSplat0.6b

TripoSplat, TripoAI's single-image-to-3D-Gaussian-splat model ("Generative 3D Gaussians with Learned Density Control", SIGGRAPH 2026), repackaged for LibreYOLO. One picture of an object in, the whole object out as up to 262,144 Gaussians.

from libreyolo import LibreGen

model = LibreGen("triposplat")      # downloads this repo
model("chair.jpg").save("chair.ply")

libreyolo generate source=chair.jpg model=triposplat

This is a mirror, not a new model

The weights are VAST's. Nothing was retrained, converted or re-encoded; the three files are byte-identical to upstream and keep upstream's folder layout.

• Upstream repository: VAST-AI/TripoSplat
• Upstream revision: 56a96e603204ec410c4da60c13ea4fa09a2169a9
• Code: VAST-AI-Research/TripoSplat at d8db9e018b413dd9c4a9fe22463781bf98e8e68d
• Paper: arXiv:2605.16355

File	Network	SHA-256
diffusion_models/triposplat_fp16.safetensors	the 370M rectified-flow denoiser	c870b97ac1d6bc9177608a5ec625e19ef9f3c5019aa68f64b0fb7803abcd6d20
vae/triposplat_vae_decoder_fp16.safetensors	the octree and Gaussian decoders (288M)	ed0d0c3d43b599e326845d0ec70f3cf77be9a55e2d97627ac3b34d2830763cc8
background_removal/birefnet.safetensors	BiRefNet (Swin-L) for background removal	9ab37426bf4de0567af6b5d21b16151357149139362e6e8992021b8ce356a154

Upstream's DINOv3 and FLUX.2 VAE copies are not mirrored. LibreYOLO reads the DINOv3 ViT-H+/16 encoder from LibreYOLO/LibreDINOv3vith16plus (Meta's DINOv3 License, which restricts some uses) and the FLUX.2 VAE from black-forest-labs/FLUX.2-klein-4B (Apache-2.0).

License

MIT, as published by VAST. See LICENSE and NOTICE.