HD Porn Finetune (SD)
Download the training data for the trigger list.
It also contains good and bad example images showing what this model can produce.
Finetuned to 0.03 Loss
Extremely high success rate (90%+) on anatomy placement for females (Vulva, Anus)
Extremely high success rate for Oral Sex (Blowjob)
Fair success rate on Anal Sex and Sex
Fully fined tuned across multiple subjects.
Trained on top 100 Adult Film stars
Trained on nature pictures
Limited training on 3D and Anime
Using "--fp32-text-enc" is recommended by not required.
HD Porn Finetune (Toon Ed.)
Its not Pony but all the images shown in the main image can be replicated seed to seed except the cover which is a image to image with ERSGAN2x
Description
FAQ
Comments (11)
This is honestly fairly impressive for 1.5, but how's it fair with SDXL/Pony?
It isn't going to match PONY for generation - unless your using it for a face replacement with yolo or adetialer for speed
The toon version only seems to produce realistic images. Is that the correct version of the model?
I didn't wipe the hole model with cartoon, so you will need to use a trigger word - you can load the images for workflow
@Felldude I see.
I tried using keywords in the examples and couldn't get it to do anything, but it may be that my workflow is screwing something up. Thank you for the response.
@allie You should be able to load the images a directly reproduce them, beyond that I am not sure. Your not replacing the CLIP are you
you can do illustrious version?
I have done some work with illustrious but the clip has some very baked in artist and characters. The dataset is public but would be a nightmare to clean up
It's a pity that this model doesn't work in a cartoon style, so to speak, "out of the box". Unfortunately, my test set didn't include the right triggers. Anyway, the results are not bad in their NSFW direction.
Thank you, as to your point that poses an interesting question, one that I am still not sure on. Do these models really learn to cross concept, or is the fine tuning just making it easier for it to retrieve one of the billions, or for new models trillons of images it was trained on.
@Felldude Anyway, great thanks for this work! (And for tip). I understand, that making such thing, it's not just "weekend park walk".
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.