Introducing Venus XL
The new champion of female nudes for SDXL
This model is based on a dataset of around 5000 nude images that were MANUALLY selected for aesthetic quality, scaled, cropped, captioned, and tagged with booru tags too. Did I mention this was all done MANUALLY!? Hundreds (or thousands) of hours were invested in creating the dataset!
This means you can prompt with only captions, or only booru tags. Yes, "1girl, completely nude, large breasts," will give great results. So will using normal English sentences! It retains the ability to create any art style, and seems to have only improved SDXL.
In addition to nudity, both hands and faces seem to be improved over SDXL base, but it doesn't hurt to use Face Restore for improved quality. Using the SDXL refiner will improve faces, but also tends to put clothes back on.
NOTE: The best VAE for this model is BUILT IT, and is the SDXL 0.9 VAE. You don't need to use any external VAE.
ANOTHER NOTE: To further improve the results, I merged in weights from other models. This final result was made with 80% Venus base (my trained model) plus 5% MiaModel, plus 15% Osorubeshi Alpha.
Edit: Osorubeshi percentage was low lowered in version 1.1 as it made skin look plastic.
Description
Reduced the amount of "Osorubeshi Alpha" in the model to combat plastic-looking skin textures. Use slightly lower CFG if results look waxy or plastic.
FAQ
Comments (8)
Not good, easy to cartoon
Also my experience. Tried with several prompts in very different prompting-styles and always get something that reminds me of 1.5 base model...
I am increasing the dataset and will be training again from scratch in the next few weeks. I still think it outperforms other realistic SDXL models for nudes, but not as reliable as newest 1.5 models.
@vegakh from my tests it underperforms in every regard even compared to SDXL base... Looking forward for the next attempt; it does seem to work in some way for you, that's promising enough.
@traxx2012 I have tried it as well and agree it is very poor on it's own, however, the data in it when combined with better models produces some very good results. Better than those original models on their own. So I say good work and keep going. Add more high quality data into the more advanced models is what helps them all get better. I have a few test models that I mixed this one into and they look great. My original AcornXL is pretty good, so I don't know when I might release the next version. I don't release, just to release, so only when I am convinced I have significantly improved on my original will it come out. This model will definitely be part of the mix though.
Keep em coming good sir, as a stand-alone checkpoint it needs more work, but it fixes the other checkpoint's lack of good nudity.
Hopefully, it gets better, so I can use it on its own
I really like this model and get good results.
Good mix for nsfw loras