>>55224947TIs, for now.
I am comparing a lot still
both the same runs one time with subject vs subject_filewords (the txt subtitles)
and then different training schedules/ cross models (Original NAI, Anything V3, Nothing V2.3) and with different VAEs.
I try to stay close to the original representation but it's difficult with more recent models. For example this Misty has been trained on original art but the model enforces the "newer anime style" regardless. It may work better with a LORA, I will have to run experiment. I can capture the essence better using older models like Anything V3. This is likely due to the Curse of Recursion (
https://arxiv.org/abs/2305.17493) where training on AI images makes the model forget the original probability distr.
I am still automatically scraping a large anime DB (from anime series directly) to eventually train a better anime model based on SD v1.5 (instead of using anything that was trained on the NAI leak from over a year ago)
