Common Canvas: Open Diffusion Models Trained on Creative-Commons Images

Prompt SD2 CommonCanvas-S-C CommonCanvas-S-NC CommonCanvas-L-NC an oil painting of a tall ship sailing through a field of wheat at sunset Figure 1 . We achieve comparable performance to public Stable Diffusion 2 (SD2), using entirely Creative-Commons images and a synthetic captioning approach that requires only <3% of the amount of the data used to train previous models. We include results for two CommonCanvas architectures, small (S) and large (L), and two CC-image datasets, commercial (C) and non-commercial (NC).