I’ve never made a LORA before. 50 images of cocks burned to a crisp at 1e-4 for 7000 steps. Catastrophic forgetting of everything but penis.
v2.0: The dataset has been expanded to 400 images including men, women, and anime/hentai. LoRA was retrained using prodigy for 24600 steps.
Description
After some testing, I've found that using replacement code words actually has a quite negative effect on results overall. This no longer uses keywords, and is prompted with natural language without any trigger. However, the captions are a bit formal, so it knows better what a 'penis' is rather than a 'cock', and so on.
Expanded the dataset to 400 images by adding women and anime/hentai images. Retrained using prodigy for 24600 steps.