The training dataset consists of screenshots from her re-debut (2.0 model) and 3d showcases. Given the nature of the dataset, you might have to do some further upscaling when generating full body shots. The lora works best for upper-body shots, at least on weights close to 1. On rare occasions it will generate the wrong hair color, just add it as a negative prompt if it ever happens.

Some details about training:

• dataset_count = 66

• optimizer = CAME

• learning_rate = 0.0004

• learning_rate_scheduler = cosine

• epochs = 11

• steps = 726

• lora_rank = 32

• lora_alpha = 1

• batch_size = 1