A LoRA model trained using Kohya. Trigger word: pekomama
Tends to be sticky at 1.0 strength, I recommend you use 0.5 strength to get other outfits.
Technical Details:
Resolution: 1024px
Network: LoRA (32/16 dim/alpha)
Steps: 2525
Optimizer: Adafactor
Learning Rates: 1e-4 (UNet), 5e-5 (Text Encoder)
Features: Bucketing, caption shuffle, flip augmentation
SNR Gamma: 5, Noise Offset: 0.1
Schedule: Cosine with 3 restart cycles
Use "pekomama" in prompts.