LibreTRELLIS2-4b

The geometry networks of TRELLIS.2-4B, Microsoft's image-to-3D model ("Native and Compact Structured Latents for 3D Generation", 2025), repackaged for LibreYOLO. One picture of an object in, the object's shape out as a triangle mesh.

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

from libreyolo import LibreGen

model = LibreGen("trellis2")        # downloads this repo
model("chair.png").save("chair.obj")

libreyolo generate source=chair.png model=trellis2

This is a mirror, not a new model

The TRELLIS.2 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.2-4B
• Upstream revision: af44b45f2e35a493886929c6d786e563ec68364d
• Code: microsoft/TRELLIS.2
• Paper: arXiv:2512.14692

File	Network
ckpts/ss_flow_img_dit_1_3B_64_bf16	sparse-structure flow transformer
ckpts/slat_flow_img2shape_dit_1_3B_512_bf16	shape-latent flow transformer, 512
ckpts/slat_flow_img2shape_dit_1_3B_1024_bf16	shape-latent flow transformer, 1024 and 1536
ckpts/shape_dec_next_dc_f16c32_fp16	shape decoder (flexible dual grid)

Not mirrored: upstream's texture networks and the shape and texture encoders, which LibreYOLO does not use. The sparse-structure decoder that TRELLIS.2 reuses from TRELLIS and the BiRefNet background-removal weights load from LibreYOLO/LibreTRELLISl; the DINOv3 ViT-L/16 image encoder from LibreYOLO/LibreDINOv3vitl16 under Meta's DINOv3 License. SHA-256 of every file is listed in NOTICE.

License

MIT, as published by Microsoft. See LICENSE and NOTICE. The DINOv3 encoder the model reads its input through is not MIT; see its mirror's card.