#nástroj tipy na trénování vlastních LoRA modelů
Reddit - Dive into anythinghttps://www.reddit.com/r/StableDiffusion/comments/13dh7ql/after_training_50_lora_models_here_is_what_i/Style Training :
- use 30-100 images (avoid same subject, avoid big difference in style)
- good captioning (better caption manually instead of BLIP) with alphanumeric trigger words (styl3name).
- use pre-existing style keywords (i.e. comic, icon, sketch)
- caption formula styl3name, comic, a woman in white dress
- train with a model that can already produce a close looking style that you are trying to acheive.
- avoid stablediffusion base model beacause it is too diverse and we want to remain specific
Person/Character Training:
- use 30-100 images (atleast 20 closeups and 10 body shots)
- face from different angles, body in different clothing and in different lighting but not too much diffrence, avoid pics with eye makeup
- good captioning (better caption manually instead of BLIP) with alphanumeric trigger words (ch9ractername)
- avoid deep captioning like "a 25 year woman in pink printed tshirt and blue ripped denim striped jeans, gold earing, ruby necklace"
- caption formula ch9ractername, a woman in pink tshirt and blue jeans
- for real person, train on RealisticVision model, Lora trained on RealisticVision works with most of the models
- for character training use train with a model that can already produce a close looking character (i.e. for anime i will prefer anythinv3)
- avoid stablediffusion base model beacause it is too diverse and we want to remain specific