Model description

What was improved compared to v4

• v4 issues: Looked very silky/smooth and felt slightly underfit.
• New training:
  • Updated dataset with less perfect, torn, and filthy day-to-day clothing in a wide variety of real-life situations.
  • Much more detailed skin textures.
  • Used 5-Preview Detailer Hook Provider workflow to enhance fine details.
  • Higher network rank + slightly more conservative learning rate so the LoRA actually learns the details properly.

Results

• Better, more natural skin quality
• Much closer to the character's original facial identity
• More realistic and varied clothing (including imperfect/torn/filthy everyday looks when prompted)
• Overall more balanced and better output

Trigger words

Download model