1.0：
Due to a personal operational error, remnants of other artists' styles remain, which has significantly weakened this artist's style.
Because the training dataset uses censored examples, the generated private parts will have light mosaics. If you mind this, please do not use it (or wait for me to train an uncensored version).
In the future, I will mainly use uncensored high-resolution training datasets, This should be able to remove the light mosaics.
I set a relatively high training resolution, so it still performs well even under high step counts. (Example image: Esdeath)
I'm not very satisfied with the results of this training, especially the face part; so please look forward to v2.0！
2.0：
he training set used uncensored ultra-high-resolution original images. Training ran for approximately 12 hours and more than 11,000 steps.
There are no mosaics remaining.
Hands sometimes exhibit quite noticeable deformations. It is recommended to use Hand ADetailer or other detail-improving LoRAs.
Because the training set consists of ultra-high-resolution images combined with up to more than 11,000 training steps, this LoRA should readily accept and perform well with higher-resolution image generation.
Similarly, simply increasing the number of sampling steps can bring the detail and art style reproduction closer to the original. 40–50 steps already yield good results (though I personally still find them somewhat stiff). Reaching around 100+ steps generally achieves approximately 80% of the original artwork’s style.
Specific prompts may cause over-merging of training set content, but I consider this to be an acceptable minor flaw.
Regarding my generation settings for v2.0
sample：er_sde
scheduler：sample
cfg：3~7
steps：30~50 or ｍore steps（The higher the steps, the better the art style restoration, but it relatively increases the likelihood of some noise and artifacts appearing）