Train using 35 images from Ai Shungaya (https://www.pixiv.net/en/users/114387853)
This is the 4th and successful attempt at replicating the style because the training somehow has high diminishing return for body proportion & style at 100 images. Watermarks are cropped out manually.
After testing using checkpoints on civitai that I think is ok, VXP IL Base Model (NSFW alert!) is the most compatible so far, while WAI v17 is 50/50.
VXP IL Base Model (NSFW alert!) - 9 steps, CFG 4, Euler A, Liquid9745VAE,
Weight: 1.0-1.3. Euler or Euler a is fine, but preferably Euler a, also depending on checkpoints. How anime/ADetailer the face is looking depends on the random shadow, but using Euler a can get consistent face.
Recommended to use detail booster lora & saturated VAE.
May add "(saturated:1.3), (high contrast:1.3)" but I doubt it does anything extra.
Or
Any other better checkpoints or control settings of your own preference.
Enjoy!
![莉缇娅 [极乐之石]-艾尔之光（Lithia [Gembliss]-Elsword） - AI Model cover image](https://images-ng.pixai.art/images/thumb/2f81ead3-c577-4fd0-8bb3-1e088017dd35)