What is this?
The base SD3.5 Large model lacks training data for nudity, which is a hinderance even when doing SFW prompts.
To solve this , this LoRa serves "fill in the blanks" in the SD3.5 for female nudity and anatomy.
Best usage for this LoRa:
Pick an image online (something like a magazine page will work well) run it through JoyCaption , and use the prompt.
//---//
Training method:
This LoRa was trained on SD3.5 Large using fp8 precision.
Training images are 45 images of naked female anime characters on a single color background.
As you can see in this chart, even with different characters and perpectives , the loss rate is low at ~0.09.
//---//
Here is an article I'm writing on my LoRa creation process with SD3.5L : https://tensor.art/articles/794137808332900353
//---//
Prompts for training images are created using Joycaptions: https://huggingface.co/spaces/fancyfeast/joy-caption-alpha-one With the keyword 'naytlayt' added close to the start of the prompt.
All training prompts were kept between 550-800 characters in length. The reason for this limitation is to keep the text prompt within a single 256 batch encoding for the T5 encoder.
To quote Stability AI:
"While this model (SD3.5) can handle long prompts, you may observe artifacts on the edge of generations when T5 tokens go over 256. Pay attention to the token limits when using this model in your workflow, and shortern prompts if artifacts becomes too obvious."
Source: https://huggingface.co/ckpt/stable-diffusion-3.5-medium
//--//
Feel free to use this LoRa however you like. You may merge this LoRa into your own LoRa project and/or model without crediting me.
