Recommended strength: (1.40)
Recommended words on prompt (no trigger word): Erect Penis, Erect penises (for more than one person).
The model was trained in three stages:
Musubi-tuner:
V1: 13860 steps total, 6930 steps in batch size 2, 77 images, 23 hours. (undertrained)
V2: 3240 steps in total, 1620 steps in batch size 2, 54 images, 5 hours.
V3: 3240 steps in total, 1620 steps in batch size 2, 54 images (same dataset as V2), 5 hours.
Description
FAQ
Comments (8)
Thanks, it works great! Any chance you could make a version that works with Qwen Edit? I tried it and it didn't work (not surprising I guess).
works for me
Thank you! I don't really know how to train a LoRA on Qwen image Edit, but I do know an inpainting method for Qwen that works pretty well with the LoRA, hope this one works for you: https://www.reddit.com/r/StableDiffusion/comments/1nhlzjq/qwen_inpaint_preserve_output_quality/
@ms735 hmm, I'll try again but definitely didn't work for me in Qwen Edit....
@JKf_05thanks!!
Sorry I gotta ask : what's the thing with that byvalyi dude's images below? Most of them seem to be missing the part this LoRA is supposed to help with.
Are the images like an indirect jab to the creator that the LoRA doesn't work?
Did he just decide there were too many penises in this specific photo section and they had to be balanced with vaginas?
I don't mind at all, there weren't any vaginas in my dataset, but maybe the downgrade in quality that this LoRA achieved is what they like, lol. Besides, they are using the Jib Mix Checkpoint, wich may not work well with this LoRA since it was trained with the Base model.
They're autodetected. Most likely case is that the person tried using this LoRA for something else, then changed focus and generated babes instead, forgetting to turn off the LoRA, then posted them with metadata that includes the LoRA and it's gotten auto-detected and injected into the gallery down below. Since there is very few works featuring this LoRA on the site — the presence of these is apparent, but would be drowned in better works if many were posted.