Improved version of the SD 3.5 Medium Node Woman Lora. It produces a more natural looking nude female body in multiple positions. Breasts and nipples are far more accurate, and the pubic area is far more accurate.
In some prompts you'll get a good photo right of the bat in others it takes a very descriptive prompt. I originally attempted to train a separate Lora for more accurate vaginal area, but that took forever to converge and was kind of limited.
I decided to merge the training set for that Lora into A slightly reduced set of the original 1.0 version's models. That helped but it still has some issues. So I added a few more poses of the female body in both closeup, head to toe, legs spread, etc. It took a set of 376 images to get a good output.
I also decided to re-caption all the images with Joycaption-alpha-2 with some edits. I think that helped a lot.
The current trigger word is "nudewoman", but the more detailed the description the better. However due to the large number of images trained against, you don't typically need to include the trigger word, at least in ComfyUI.
I found in testing first try it without the trigger word and if that doesn't look good try it with the trigger.
Strength works from 0.5 to 1.5, but the best range seems to be 0.75 to 1.15.
Description
This is the initial version trained on 187 images for 27 epochs. It was trained in ai-toolkit using Joycaption captions manually edited, so it should work with verbose captions. I tested it in ComfyUI using strength 1.0, 40 steps, cfg 5.5 with DPM++2M with the sgm uniform scheduler. It also does well with Euler and Heun using either normal or simple scheduler.