Testing out training styles on Illustrious-XL.
Here's what the different versions are:
003 - Trained without any artist tags. Works well alone but has little effect alongside any artist tokens.
006 - Trained using Ishikei as a token either at the front of the prompt or behind character names. Works well when using Ishikei as an activation token, but not other artist tokens.
007 - Trained using the top 1000~ most prolific artists on Danbooru as wildcards at the front of the caption or behind character names. Works well no matter what.
019 - Trained using the top 1000~ most prolific artists and around 400 copyright tokens from Danbooru, as well as a large number of hyperparameter changes. Should be more resilient when used alongside copyright tokens while hopefully looking a little nicer.
These same approaches also apply to PonyXL and probably any other model with a similar architecture, but with PonyXL the top 1000 copyright tokens from Danbooru/rule34 should be applied in place of artists in order to reproduce this effect.