These are Kaname Madoka (uniform and transformed) and Ultimate Madoka from Puella Magi Madoka Magica. Coincidentally, the bow and arrow and school bag are also perfectly reproduced. There are 20 sample images.

There are 102 training images in total: 27 uniform images, 48 transformed images, and 27 Ultimate images. For some reason, even when setting the strength to 0.1, the reproduction quality, including the fine details, remains the same. This is the first time I've seen this, and it's puzzling. The results were the same when I created Lora in Ultimate or transformed form alone.

Added 10/19/2025

We have discovered that the phenomenon of overfitting when eye color changes or when the intensity is high is unique to SeaArt's previous mode. This phenomenon was not observed in SeaArt's new mode or Pixai. Furthermore, by using SeaArt's new generation mode quality, we can generate HD images that significantly exceed the quality of the sample images.

The recommended strength for uniform and Ultimate is 0.2-0.5. The recommended strength for transformed form only is 0.3-0.2. While there are no problems with reproduction quality even at 0.1, sometimes one out of four images tends to look like it was painted in a gal game style. To change the hairstyle, you need to include a hair down prompt, otherwise the image will remain in twin tails or a two-sided updo. Even when including a hair down prompt, there are sometimes images with the same hairstyle.

At 0.7 or higher, Lora's influence is so strong that the drawing style feels overtrained, and after transformation, the eye color often changes to purple or orange. Eye color can change even in uniforms. This tendency is particularly noticeable after transformation, perhaps due to the large number of training images. At 0.3 or lower, the eye color barely changes.

The behavior of the eyes seems to vary depending on the model; Hoseki - LustrousMix [IllustriousXL] and WAI-NSFW-illustrious-SDXL (V13) barely changed eye color even at 0.8. WAI-NSFW-illustrious-SDXL V14 and V13-9 changed eye color. I felt that V13 was poorly rendered holding a bow and arrow. At higher intensities, the eye color changes more easily as the character moves, such as in combat stances. Hoseki - LustrousMix [IllustriousXL] produces slightly more realistic anime drawings than WAI-NSFW-illustrious-SDXL.