Style-LoRA of the artist Nagi (ナギ) trained with his illustrations, manga images, and miscellaneous images on Pixiv and other places. Full credit to the artist, who has not been active for years. His works left an impression in the community and inspired a lot of creations of Villainess (敵女) and Ryona (リョナ) themed creations.
This first version is still in experimental phase, with rough tagging and compromised performance. I'm uploading it as a PLACEHOLDER to test its results. Will train with improved stability and more versatility in the future. Would appreciate any reviews below -- show me your works!
Training dataset: 303 from artists post works
Settings
Trigger word: nag1art
LoRA weight: 0.7~0.8
Checkpoint: WAI-NSFW-illustrious-SDXL
It's recommended you use this LoRA with Illustrious-based models. However, you can download the model and use it with any other Pony or SDXL model you want, while the fidelity of the character might be compromised.
Resolution: 832x1216 or 1024x1360
These are the dimensions in the model version features of most Illustrious checkpoint models.
Sampling method: Euler a
CFG scale: 7~12
Sampling steps: 25~30
Hires. fix (optional): R-ESRGAN 4x+ Anime6B
Steps: 20+
Denoising strength: 0.5
Follow me on: