darkhatboy style | Anima LoRa

style based on darkhatboy's art.

spent over a month training many Loras and experimenting on Anima to figure out how to train it. ended up having to learn diffusion pipe and train it there because i was having a lot of issues with prompt adherence on kohya's scripts. hence why there's no metadata on my lora file.

i did this all in free hardware on colab and eventually lightning studio which is why it took so long.

but finally i have a lora style that passes all my prompt test.

generation recomendations from my test runs:

Steps: 30

cfg: 4-5 i usually go with 4 to 4.5 depending on the prompt.

sampler: er_sde, and euler_a. use er_sde for this style

scheduler: simple

positive prompt: masterpiece, best quality, absurdres, [insert artist name here], [your tags]

negative prompt: worst quality, low quality, score_1, score_2, score_3, lowres, bad anatomy, bad hands, bad feet, bad arm, bad leg, bad neck, bad perspective, bad proportions, extra digits, fewer digits, wrong hand, wrong foot, artist name, signature, watermark, [you can optionally add english text if you dont want text]

and adetailer when the face is far away/in general.