SUGGESTED RESOLUTIONS: 1024x1024, 832x1216, 1216x832, etc. Avoid using resolutions that are too small or excessively large.
I have not used this yet but this information is from the creator.
The model was soon killed off after I downloaded it. However, I was able to recover information. The model was uploaded onto civitai by user Bubble75.
Euler a, cfg 7, and 30 steps has been working well.
Check the gallery for more details on styles. The model is highly Lora compatible and I would actually recommend that you use style loras as they seem to stabilize style consistency.
Pos:
general tag -- describe what you want quality tag -- best quality,high quality,normal quality,low quality,worst quality Neg: Check gallery for examples as the creator suggests none.
17 Artist tags recognizable by the model:
atdan, misaka 12003-gou, qing wu, ru zhai, swkl:d, teddy (khanshin), vardan, houkisei, icecake, kamue, nekoda (maoda), ririko (fhnngririko), shanguier, tansuan (ensj3875), wd (1106592840), qizhu
You are encouraged to repeat the artist's name. This makes the style more prominent.
For instance: qizhu, By quizhu, by artist quizhu, in quizhu style ...
Author's Note：
This model was trained using a dataset of about 5,000 images, with each artist's work equally represented (~300 images each)
It excels at generating detailed feet. (Pretty sure I disagree with his statement) You can look up the artists' works to see the quality.
The training used the Lokr algorithm from the Lycoris project, running for 10 epochs on a 4070 Ti Super for approximately 10 hours. The result is a 78.2MB Lokr file, merged back into PDXL with a weight of 1. So, this model is entirely trained and not merged with others post-training.
There is no difference between v1.0.1 and v1