Thanks for the training :D. Just a thought about the tags:

- "positioned from thighs up" should be tagged: cowboy shot

- "partially hidden by XX" should be tagged: transparent

- "holding XX with both hands" should be tagged: holding, etc.

You seem to have tagged an illustrious dataset with tons of words and phrases than mean nothing to the base model. Many of your words are not existing tags, so they just add noise, which doesn't help the model to make sense of pixels during training.

"holding, transparent, swim ring" would probably deliver a more efficient training. For Illustrious or any other model based on booru tags (Pony, NoobAI, etc.)

Always check danbooru to know if tags you caption actually exist ;)

Or tag using BooruDatasetTagManager, loading a danbooru tags from around the time the model was trained; because tags evolve, are changed, removed, etc. Just put the .csv in the \BooruDatasetTagManager\Tags\ folder.

Illustrious for example was released in sept. 2024, and they used a danbooru list of tags from 2023 up to early 2024 (april 2024).

Good luck with your trainings ;)