YUN Character LoRA – Revised Version

This version was retrained with a redesigned dataset to improve outfit flexibility while preserving YUN’s core identity.

Compared with the previous version, the training images were rebalanced across multiple outfits, viewing angles, and portrait/full-body compositions. YUN’s fixed character traits, such as her hairstyle, eyes, horns, and facial features, were kept consistent throughout the dataset, while clothing, camera angle, and framing were explicitly treated as variable attributes.

The main goal of this update is to reduce the tendency of the LoRA to force YUN’s default outfit when alternative clothing is requested.

Main improvements:

• Better separation between character identity and clothing

• More balanced outfit distribution in the training dataset

• Improved compatibility with alternative outfits

• Stable character features across front, side, back, and three-quarter views

• Improved consistency between full-body and upper-body generations

Epoch 7 was selected as the main release version because it showed the best overall balance between character consistency and outfit flexibility during testing.

Trigger word: yun