LibreTRELLISl

TRELLIS-image-large, Microsoft's image-to-3D model ("Structured 3D Latents for Scalable and Versatile 3D Generation", CVPR 2025), repackaged for LibreYOLO. One picture of an object in, the whole object out as 3D Gaussian splats.

LibreYOLO runs it in plain PyTorch, without spconv, xformers or flash-attn, so the same model runs on NVIDIA GPUs, Apple silicon (MPS) and CPU.

from libreyolo import LibreGen

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

libreyolo generate source=chair.jpg model=trellis

This is a mirror, not a new model

The TRELLIS weights are Microsoft's. Nothing was retrained, converted or re-encoded; every file under ckpts/ and pipeline.json is byte-identical to upstream and keeps upstream's layout.

• Upstream repository: microsoft/TRELLIS-image-large
• Upstream revision: 25e0d31ffbebe4b5a97464dd851910efc3002d96
• Code: microsoft/TRELLIS
• Paper: arXiv:2412.01506

File	Network
ckpts/ss_flow_img_dit_L_16l8_fp16	sparse-structure flow transformer
ckpts/ss_dec_conv3d_16l8_fp16	sparse-structure decoder
ckpts/slat_flow_img_dit_L_64l8p2_fp16	structured-latent flow transformer
ckpts/slat_dec_gs_swin8_B_64l8gs32_fp16	Gaussian decoder
ckpts/slat_dec_mesh_swin8_B_64l8m256c_fp16	mesh decoder (not used by LibreYOLO yet)
birefnet.safetensors	BiRefNet (Swin-L) background removal, MIT, by Peng Zheng et al.

The radiance-field decoder and the structured-latent encoder of the upstream repo are not mirrored. LibreYOLO reads the DINOv2 image encoder from facebook/dinov2-with-registers-large (Apache-2.0). SHA-256 of every file is listed in NOTICE.

License

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