1. Label: When training a flux lora of a real person, label(including tag and caption) is not necessary. Flux itself has strong character learning ability. Label is only necessary when training styles, accessories, clothing, concepts, etc.
2. Max train steps: 1200-1500 is sufficient, more steps will not bring better results.
My training parameters:
width & height 768*768,
batch size 1,
optimizer adamw8bit,
lr scheduler cosine with restarts,
learning rate 4e-4,
network dim 64,
network alpha 32,
3. Datasets: 15-20 photos are enough, more photos will not bring better results. There is a little trick to choosing photos. You must choose one that looks like the person. If you cannot even identify whether the person in the photo is herself, delete the photo from your datasets.
