LibreCRM

Use restrictions apply. The two diffusion checkpoints descend from Stable Diffusion 2.1 and are distributed under the CreativeML OpenRAIL++-M License (see LICENSE). Its paragraph 5 and Attachment A forbid some uses (among them unlawful content, harming minors, defamation, harassment, discrimination and medical advice); by using the model you agree to them, and you must pass them on to anyone you distribute it to.

CRM ("CRM: Single Image to 3D Textured Mesh with Convolutional Reconstruction Model", ECCV 2024) by Zhengyi Wang et al. (Tsinghua University), repackaged for LibreYOLO. From one picture of an object, two multi-view diffusion models generate six views and their canonical coordinate maps, a convolutional UNet turns them into a triplane, and FlexiCubes extracts a coloured mesh.

LibreYOLO runs it in plain PyTorch, without xformers, nvdiffrast or a GPL mesh library, so the same model runs on NVIDIA GPUs, Apple silicon (MPS) and CPU.

from libreyolo import LibreGen

model = LibreGen("crm")                  # downloads this repo
model("chair.png").save("chair.obj")     # vertex colours
model("chair.png", texture=True).save("chair.obj")   # UV texture (pip install xatlas)

libreyolo generate source=chair.png model=crm

This is a mirror, not a new model

The weights are the authors'. Nothing was retrained, converted or re-encoded; the three .pth files are byte-identical to upstream.

• Upstream repository: Zhengyi/CRM
• Upstream revision: 36283fd567b9a3120a166dfe2f2e0965d96261b0
• Code: thu-ml/CRM
• Paper: arXiv:2403.05034

File	Component
pixel-diffusion.pth	Multi-view diffusion: six views from the input image (UNet, SD 2.1 VAE, OpenCLIP ViT-H/14 towers)
ccm-diffusion.pth	Multi-view diffusion: canonical coordinate maps of the six views
CRM.pth	Convolutional reconstruction model: triplane UNet, SDF, colour and FlexiCubes-weight MLPs

License

CRM's authors tag these weights MIT. The diffusion checkpoints are a fine-tune of ImageDream and MVDream, themselves fine-tuned from Stable Diffusion 2.1 base, whose CreativeML OpenRAIL++-M License requires every derivative to carry its use-based restrictions, so this mirror ships that license verbatim (Stability AI, dated November 24, 2022) together with CRM's MIT grant. See NOTICE for attribution and SHA-256 of every file. This is LibreYOLO's reading of the terms, not a statement by the upstream authors.

Citation

@article{wang2024crm,
  title={CRM: Single Image to 3D Textured Mesh with Convolutional Reconstruction Model},
  author={Zhengyi Wang and Yikai Wang and Yifei Chen and Chendong Xiang and Shuo Chen and Dajiang Yu and Chongxuan Li and Hang Su and Jun Zhu},
  journal={arXiv preprint arXiv:2403.05034},
  year={2024}
}