Unofficial mirror to build hype on the release of the Common Canvas line of models by MosaicML.

( I am not the creator, don't credit me unless you want to, lol)

Common Canvas models are trained on CC and Public domain works using the same architectures as SD2 and SDXL for the Small and XL types, respectively, using the unique training approach of PAG, or Perturbed Attention Guidance, allowing for significantly smaller datasets that achieve similar quality to models that until now only large companies could afford to train. There are two main models, Common Canvas XL Commercial, and Common Canvas Small Commercial, along with two even larger, but noncommercial, trained on CC-NC, XL NC and S NC models.

Additional files and resources are available at the Hugging Face official repo, including access to the entire catalogue of training data.

Link to Hugging Face demo page:

https://huggingface.co/spaces/common-canvas/CommonCanvas

Link to the original paper:

https://arxiv.org/pdf/2310.16825

Link to the Github repo used for training, which is also recommended for further finetuning:

https://github.com/mosaicml/diffusion

Licenses are Open RAIL++-M License, for CCXL, and Creative Commons Attribution Share Alike 4.0, for CCS.

(I've uploaded only the Commercial models here.)

(I've selected the Civitai options to try and match, otherwise would just leave open on everything.)