Training new characters?

#2
by EndlessHours - opened

Hey, this is awesome! If possible, I would like to train some additional characters, such as Kit Fisto, but I am fairly new to SD. Are you able to share how you went about training this model?

Sorry for late response. You can get the general idea from my reddit comment:

I used this repository, and mostly followed these instructions.
But in the concepts JSON each concept's instance_prompt matches class_prompt (e.g. both are "clone wars style", "anakin skywalker", etc.) and instance_data_dir matches class_data_dir. 20 images per concept.
Actually, regularization images weren't used at all. But I wanted more images for style, so I split style images to different directories, so there are 10 concepts with same "clone wars style" but different directories.

Then, the model was trained with these parameters:

learning_rate=1e-6
lr_scheduler="polynomial"
max_train_steps=20000

And the results were fine, but I continued training with learning_rate=1e-7 and lr_scheduler="constant" for 20-k more steps. You should save each N-k steps, check for results and choose the best.

Sign up or log in to comment