Realistic BodyWriting (RBW) is for inpainting.
The model does a poor job of generating the correct text during txt2img.
The sample image was generated with URPM1.3, and then inpainted with URPM1.3-inpainting and RBW.
Recommend values when inpainting with URPM1.3-inpainting,
.7-.8 for LORA weight
prompt weight of 1-1.3
CFG scale of 5
denoising strength of .7-.85
Other values were tested but these were found to be the best ranges.
Inpainting
When inpainting, it is recommend to add one word at a time. The success rate depends on the text being inpainted.
RBW is partially case-sensitive, ie T vs t may change the case of the letters used in the image.
The size of the inpainting area and the resolution of the picture being inpainted are both very important.
The LORA is prompted with,
text "ADDTEXT" on bodypart
emoji "EMOJI" on bodypart
Common bodyparts are,
upper chest
breast
upper abdomen
stomach
navel
pubic mound
leg
ass
Notes
Text contained in the training data generates better than text that isn't.
Non-training text is still possible but has a lower success rate.
For text in the training data the success rate is (1-10%).
For text not in the training data the success rate is lower and strongly depends on the word.
The full list of texted used for training is in the LORA info (I icon in the LORA menu for auto1111).
Adding a basic description of the subject in the prompt helps. i.e. women standing naked
Source Material
The images were gathered online and the heads were removed to anonymize the images.
If anyone feels the model depicts them too accurately, I am willing to retrain the model with their images removed.
Conversely, is anyone is interested in providing custom images please contact me.
Version 2
The model was retrained many times with different parameters and tested with different epochs.
After determining the optimal parameters the dataset was expanded.
Ultimately, the version released is the best so far.
If the LORA can be improved a version 2 will be released.
Description
FAQ
Comments (9)
Can you list text and emoji contained in the training data, please?
All of the captioning data is contained in the LORA information, are you unable to access it?
@StableNudeAI Hello there, it maybe a dumb question but, where do I find LORA information on SD 1.5? Thanks for your help in advance.
No dumb questions. These little details can be hard to find. :)
The LORA information in auto1111 is the "I" icon in the corner of each LORA when you select it. If you are missing fonts it may also be a rectangle.
Tried several times with no good result. Please add some quality generated image with generation data.
Are you trying to generate an image using img2txt or are you inpainting?
The LORA isn't able to generate images of any quality using img2txt, its intended use is inpainting.
你好，我想请问一下它能输出中文吗？
I translated this as "Hello, I would like to ask if it can output Chinese?".
The model has not been tested it with Chinese characters. As it wasn't trained with Chinese bodywriting it is unlikely to work.
该模型尚未使用汉字进行测试。由于它没有接受过中国人体写作的训练，因此不太可能奏效。
@StableNudeAI 好吧，谢谢你的答复~
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.