This Checkpoint is under construction!
Please read this post π
I wrote an article about this model and the new strategy for captioning, click here to check it out.
All sample images are zero-shot generated
I will always try to avoid prompts that are similar to those used in the dataset. This way, I can show you the progression and quality of learning, making it clear that the model is learning the concepts and not just memorizing the images.
Even though no faces were blurred or removed from the images in dataset, you'll still see that it will always generate random faces (you won't see Mia Khalifa or Kendra Lust in the generated images π)
Prompt upsampling
The sample images were not upsampled; the prompt was passed directly to the main text encoder of Qwen 2.1 Turbo. The clever approach of using a dedicated text encoder solely for upsampling complicates the training process somewhat, as you must use captions that resemble the output the PUP (Prompt Upsampler) will generate. I decided to perform a fine-tuning run that works for the base model while remaining compatible with any PUP. I have been using Mistral Small and Command R Reasoning as PUPs with excellent results; I will soon post images generated using a PUP, showing both the original prompt and the detailed, upsampled prompt.
Known issues
LOSS mask images leaked into the training data; consequently, if the prompt is not understood, the model may generate a gray image.
The images generated by this model are not at an acceptable level, so don't expect extraordinary results (yet).
β = Not available yet
π΄ = Worst
π = Not good
π‘ = Almost there
π’ = Good
π£ = Perfect
Available in v5.0 dataset (initial validation)
Concepts implemented on this version:
Anus π’
Back view π’
Breasts size control (small, medium and big) π‘
Black penis π’
Blowjob π‘
Cum in mouth π
Cum on face π‘
Pussies π’
Doggy-style (dedicated) π‘
Fingering β
General sex positions (Missionary, Doggy, Spreading, POV, etc) π‘
Groping β
Hard nipples (need more images) π
Naked man π’
Panties aside (need more images) π
Playing with pussy (Rubbing, touching, etc) π΄
POV (dedicated) π΄
Pussy closeup π’
Removing clothes π
See through (need more images) π’
Penis control (flaccid, hard, small, medium, big, thin, thick, etc) π’
Text π’
Weird things (Funny and weird images, like "stepsister stucked in washing machine" π) π
Wet clothing (need more images) π’
Female masturbation (hands inside shorts, panties, skirts, etc) π΄
African women (need more images) π’
Normalization (simple prompts for complex things) β (it is no longer necessary because of the prompt upsampler)
Concepts to be incorporated soon
Sex with panties aside (dedicated)
Big breasts (to improve breasts-size control)
Sexy clothing
E-girls
Cosplay
Amateur Selfie
Wide hips (with narrow waist)
Chubby women
Fat women
Interracial sex (dedicated)
Femboys and Tomboys (dedicated)
Public (dedicated)
MILF
Lesbian (dedicated)
Teasing
Gangbang
Footjob
Cameltoe
Bathroom (dedicated)
Asian (dedicated)
Nature (dedicated)
Gay (dedicated)
Spanking/Discipline
Items marked as "dedicated" are already present in the dataset, but more images are needed for better learning.
How about helping to make the dataset?
There are many concepts and fetishes, and I don't know all of them and may not even be able to remember most of them, so you can suggest a concept to be included in the model.
You can just suggest the concept or send images/links of the concept to my HuggingFace on LoRA Requests community.
The Checkpoint Purpose:
I'm trying to bake the best SFW + NSFW Base Model.
I focused heavily on regularization so that the NSFW wouldn't end up leaking and contaminating the native SFW side of Qwen. To do this, I needed to create a large regularization dataset, which significantly delays training, making it more expensive π₯
And of course, I'm not adding knowledge, I'm replacing knowledge that is less used to make room for NSFW.
My intention is to make this model incapable of producing anime-style images, so we will have "more space" for realistic NSFW content and prevent Qwen from forgetting how to do other things, such as writing text and editing images (called "catastrophic forgetting").
Want to help me continue?
Buy PaidCoins at Aisha-AI.com and sent them to account number 2
or
Create an account on Vast.ai using my referral code (https://cloud.vast.ai/?ref_id=212422), top up your account, use the option to transfer the balance to another user and provide my ID: 212422
Total spent on this Checkpoint:
$110 USD (π)
Description
First public version
FAQ
Comments (6)
ζεΎ ε¨nsfwηΌθΎδΈζε₯½ηζζοΌηδΈδΈͺint8ηζ¬
It is highly unlikely that I will post a quantized version, as the weight-merging process I use only functions in FP32, after which I convert back to BF16. A proper INT8 quantization involves precise mathematical operations that map the values ββoptimallyβreducing the file size by half while maintaining qualityβbut this is an advanced technique that works significantly better when you have built the entire model yourself. Consequently, if I were to release a quantized version, it would lack the same level of precision as the original FP8 version, making it not worthwhile.
Keep in mind that you can download the model in BF16 format yourself and quantize it to FP8, NVFP4, INT4, or other formats; this process does not require a supercomputer.
does it work in editing without changing faces?
Using it correctly, yes
It's not a bad start besides you saying it still needs work, though it really doesn't know cunnilingus too well, often times I get fellatio in it's place, lol. I see that for the concepts for lesbian (dedicated) to be added in, but still thought I'd bring that to your attention. I'm not exactly sure what causes it, but I'm sure it's a tagging issue. Maybe both cunnilingus and fellatio get tagged oral sex and some fellatio, depending on camera angle and deepthroat could be seen to the ai as cunnilingus, and so it confuses the two if a couple cunnilingus images are tagged as fellatio. Like, I see this often too in many models. So my guess is whatever tagger people are using, isn't tagging it proper.
You're right, and that's why I stopped training. The problem is in my dataset, or rather, the problem is in how the captions were made...
The problem is the prompt upsampler; it will always use super detailed tags, and my v4 dataset had over a thousand images with captions that leaned towards the simplistic side, and this started to make the training destroy previous concepts of the model, which is not good.
I'm already creating the v5 dataset, which has much more detailed captions to reduce catastrophic forgetting, and in addition, I also need to do masked loss on each image, so the work has almost tripled.
Finishing training Qwen 2.1 with my v4 dataset would work, but it would lose the ability to generate extremely complex images, and that would be a huge problem.