Best practices for Pony XL character LoRa training:
Dataset: Putting together a great data is of utmost importance. You don’t need too many images. 50-100 are more than adequate.
Labels: Second most important thing is good labelling. Keep them short and simple. Don’t need too many because AI understands the basic features automatically.
Trigger: Pick a simple, short and easy to use trigger word.
Repeats: This depends on the number of images. Repeats x Images should be around 350-400. E.g. 50 images means 7 or 8 repeats.
Epochs: Between 10 and 12
Network dim: 32
Network alpha: 16
Batch size: Between 2 and 4
Steps: Between 1200 and 2000
Base model: SD1.5, SDXL, or Pony
