Danbooru stores millions of tagged anime images, but it doesn't have a way to filter out NSFW content. This model was trained on 100,000 of these tags with up_score ≥ 3 for 3 epochs, so it's possible that some tags might contain NSFW descriptions.
So, just be mindful of that. Thank you for your cooperation.
For details on data preprocessing, prompt engineering, and more, please see
Fast Anime PromptGen.
I used a very similar approach to train the Danbooru version.
