First, thanks for 元影METAFILM supported the major training with 15 days 4090 24GB machine on xingluan.cn
This model is an follow up version of https://civitai.com/models/646080/openkolors-anime
Based on previous work, this model is trained on t-ponynai3 v6.0 generated images.
It was using diffusers pipeline to generation all training images and filtered some characters.
(Filtering characters is because the generated characters were not similar to original character.)
The text file was conbined with the following parts: style prefix, florence-generated text, pure generation prompt.
style prefix: 二次元动漫风格, anime artwork
florence-generated text: Using microsoft/Florence-2-large-ft MORE_DETAILED_CAPTION task to generate the description.
pure generation prompt: Removed the pony score part and keep other parts. This part included the character name for image generation.
After images generated, Kolors' MPS model was used to identify images into the following category:
below_score_2
below_score_5
below_score_10
score_10_up
score_13_up
score_15_up
After the initial training was completed, a few images was seleted as good_anatomy and worst_anatomy. Both trained seperately and merged into the model with different ratio.
It trained with blue archive, genshin, naruto, one piece and some famous characters.
Due to the training is based on AI generated Images, it only produces an approximation to original images. It wasn't really learned how the character actual look like and it is the limitation of this training method.
A more detailed character list might be released afterward. (Depends on request)
Some trained characters:
2b \(nier:automata\)
asuna \(sao\)
c.c., code geass
ram \(re:zero\)
rem \(re:zero\)
tifa lockhart
hatsune miku
lucy \(cyberpunk\)
barbara \(genshin impact\)
klee \(genshin impact\)
raiden shogun \(genshin impact\)
etc
Due to the prompt structure and training limit, only using the character name might not enough to recall the character. You might try to descript the hair color before the character like blue hair rem \(re:zero\) to recall the character. Also, if the character is mixed with other character in the same series, you might try to remove the series from positive prompt.
It is also a mistake from the training prompt structure because the character name is state in pure prompt part. Next time, I could structure the character name from the start to have a better character learning.
Suggested prompt:
Positive prompt:
character name and description, 二次元动漫风格, anime artwork, other description.
Negative prompt:
worst_anatomy, below_score_2, below_score_5, below_score_10
Additional negative prompt: horns, halo
This model also trained with some nsfw but it also isn't very strong. It should be ok to use as sfw model in general.
Please leave images and comments if you like the model. Thanks.