usage/training notes:

- trained on around 160k images from e621

- including "furry" somewhere in your prompt is beneficial

- i highly recommend using rectified-flow samplers such as this: https://github.com/KeithZ117/Comfyui-anima-sampler

- tag ordering mirrors anima and is in this order: meta, character/species, artist, general tags. tags from the anima base model still work, so feel free to mix in e6 and dan/gelbooru tags.

- tags only, natural language, or natural language + tags are all valid

- around 3200 e621 artist tags were included (i.e. @anti dev) BUT due to their wide distribution, only around 50 images per artist made it in. as a result your mileage may vary - upweighting can help, but ultimately the lack of data per artist will constrain you. artists from base anima still work, so you can fall back on those for aesthetic control.

- tags were shuffled within their categories

- tag dropout was used so you don't have to write a novel if you just want 1girl standing or 1 fat fur deepthroating a corncob.

- natural language captions were generated via gemma 31b

- training was done at 512 res for most of it, with a few 768 epochs near the end. higher res might be fucky wucky but seems usable at 1mp. try the official anima hi res boost lora if you want to push it further than 1mp.

- don't expect miracles with only 160k images. it isn't a full finetune.

- artist list here: https://huggingface.co/garaggge/160kE6AnimaLora/blob/main/artistslist.txt