Purpose
Create more accurate female photos both naked and clothed. Including reduce the AI look of skin and faces, as well as more realistic body overall both nude and clothed.
I am working on creating several LoRA's for Z-Image Base, of which this is the second.
Settings from testing:
Strength: 1.0
Samplers Tested: euler, res_multi_step
Steps: 25 or 30 is good for initial look.
40 to 50 give the best results
Prompting
There are purposefully no specific key words for this model, so as not to force the output to always nude, although there is higher chance if you use words such as breast. If you want to get nude or naked then add that.
This also means that breast sizing and shape is dependent on how Z-Image does it, although it does shift size a bit from default. From testing and based off the dataset prompts you can set the sizing a bit adding the following words: small, medium sized, medium and large.
In addition Z-Image itself determines body size based on body type descriptions such as petite, athletic, slender, toned, and curvy.
Description
This is an initial version of Nude Female bodies, that is based off of several LoRA's I attempted in Flux and Chroma, but had some issues with over training, so the dataset has been slowly improved over time, and is now producing decent output.
This is one of several loras I have been working on for a while, originally for Chroma but decided to switch to Z-Image.
Eventually I plan on having the following lora types:
Breast Sizes: Which I have previously uploaded, and is mostly trained on chest and upper body and can butcher genitalia
Nude Women: Which is this lora and gives more natural and accurate nude women full body with better anatomy.
Sex acts; Create accurate sex acts in various positions
Once these are completed the final outcome will be to take all datasets and a few others to come up with a more general purpose lycoris for sex scenes.
Training method.
This was trained on OneTrainer on a dataset of 242 images of various poses, both nude and clothed from closeup to full body at various distances at a resolution of 768.
I originally ran it through a-toolkit but had issues with converging and a lot of defects. I saw comments on how OneTrainer worked better, so ran it though OneTrainer, and it converged fairly rapidly at around epoch 35, but was very inaccurate with lots of defects when I further tested it.
So I let it train further it took a bit of time to get accurately and sharper, but then around epoch 60 it started to shift between accuracy and sharpness each epoch until around epoch 75 it started improving both and by 81 was good enough that it wasn't overcooked.
Dataset prompts
As with the previous lora, the dataset was captioned on descriptive prompts generated by JoyCaption Beta 1. I then ran the prompts in a ComfyUI workflow, through the base model and generating 4 outputs for each prompt.
If 3/4 of the outputs matched the scene the prompt was kept. If half were good it was edited and tried two additional times.
The images that didn't get to 3/4 (which was about 30%) I ran through Qwen VL. If I couldn't get them to work then I threw the photo away.