Crosspost from https://reddit.com/r/StableDiffusion/comments/104mb34/a_stable_diffusion_model_trained_on_images_of/ .
Just thought it’d be interesting to share. A lot of the kanji don’t make sense but I hadn’t really thought about how new kanji could get invented, and getting an AI to do it seems like an interesting way to get candidates.
1 comment
The model seems very poorly trained and the model creator has made the classic mistake of trying to train Japanese data with English keywords, which always leads to a bunch of nonsense garbage (like the classic ç½… = crack (cocaine)).