This Checkpoint is under construction!
Please read this post 😌
Note about v9.6:
Facing problems in results, probably I'm training a lot of concepts at the same time.
And to make matters worse, GPU prices have increased ridiculously in a short time. I had started training on an RTX 5090 paying $0.22/h and now it's over $0.50/h. The RTX 4090 is at least $0.28/h.
I appreciate the people who are sending credits on Vast.ai, but with these current GPU prices, I'll have to wait a while before continue... And if the prices don't go down, then I'll continue. Please don't forget to contribute 💜
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 😌)
The images generated by this model are not at an acceptable level, so don't expect extraordinary results (yet).
❌ = Not available yet
🔴 = Worst
🟠 = Not good
🟡 = Almost there
🟢 = Good
🟣 = Perfect
Training now 🚀
Sex normalization
Penetration control
Fingering control
Refine Base NSFW
Available in beta v9.5 (preview)
Concepts to be implemented at final version:
Anus 🟡
Back view 🟡
Breasts size control (small, medium and big) 🟡
Black penis 🟢
Blowjob 🟡
Cum in mouth 🟢
Cum on face 🟢
Pussies 🟢
Doggy-style (dedicated) 🟡
Fingering 🟡
General sex positions (Missionary, Doggy, Spreading, POV, etc) 🟡
Groping ❌
Hard nipples (need more images) ❌
Naked man 🟢
Panties aside (need more images) 🟠
Playing with pussy (Rubbing, touching, etc) 🟡
POV (dedicated) 🔴
Pussy closeup 🟢
Removing clothes 🟡
See through (need more images) ❌
Penis control (flaccid, hard, small, medium, big, thin, thick, etc) 🟢
Text 🟢
Weird things (Funny and weird images, like "stepsister stucked in washing machine" 😏) 🔴
Wet clothing (need more images) ❌
Female masturbation (hands inside shorts, panties, skirts, etc) 🟠
African women (need more images) 🟢
Normalization (simple prompts for complex things) 🟠
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 Flux. To do this, I needed to create a large regularization dataset, which significantly delays training, making it more expensive 😥
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 Flux from forgetting how to do other things, such as writing text and editing images (called "catastrophic forgetting").
So it's very likely that in the future we will have one Base Model just for realistic NSFW and another just for 2D-style NSFW.
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:
$350 USD (😭😫)
Description
When I realized I had loaded the wrong dataset, I stopped the training. It learned some things, but the training didn't finish, so this version generates some aberrations and broke some concepts.
Even so, I decided to post this version, just so you can follow the progress of the training 😌
Please help me make the best SFW + NSFW model of Civitai. More information at the end of the post.
FAQ
Comments (18)
I'm looking forward to the final version... Any chance of GGUF?
Well, once we have a stable version, then I'll quantize it in FP8 and maybe INT4
Klein derivatives seem to work out of box on city96's quantiser, like flux.
@firemanbrakeneck Yeah, I know, bro. 🙂 But I was specifically referring to the quantized "Aisha NSFW" model.
@RinatLevsha I am saying, anyone can quantise it by themselves. Doesn't need a gpu, only some disk space and enough ram to load the checkpoint.
@firemanbrakeneck Ideally, I would do the quantization myself, but I haven't found a method that does the training in native FP8; it's always done in BF16. Currently, I'm working on the official BF16 version of Flux 2 Klein, and I would really like to work with the official FP8 version of Flux 2 Klein, but when merging the training weights, there's a loss of accuracy from BF16 to FP8 (half the "data" is lost). For a quantized version to be not only small but as good as the original FP8 version of Flux 2 Klein, I need to ensure that the training weights are leading to accurate calculations of the training data, and this requires a very long training with a low learning rate 😌
I haven't posted a quantized version yet because I've only published preview versions, which aren't good enough to generate images. As soon as I have a stable version, it will be available in BF16, FP8, and maybe INT4.
You can make it yourself
i made an open source easy-to-use GUI program https://github.com/qskousen/ggufy to convert image models into gguf. it has very low RAM requirements and is pretty fast. makes it easy to make your own GGUF's instead of relying on model authors to do it, and you can customize the quantization as you like
@firemanbrakeneck Oh... okay)) I just always thought all such utilities required a high VRAM GPU. Thanks for the clarification! 🙂
@ferretduck Thank you, bro!
@RinatLevsha All the ones I have seen before did indeed require a lot of RAM, but there's no real reason for that, so I made ggufy :)
@RinatLevsha Nope, no vram at all. Theoretically, all inference can be done on cpu in ram (or even loaded piecemeal one layer at a time from disk), it's just slow as molasses so nobody does that normally (cept the fruitarians with their unified ramming). But in particular for quantification, it's not even inferencing, just rebuilding the weights in some way, a one off that takes something like an hour tops, not worth the bother with gpu limits (if the math can even be rewritten in cuda, that is).
@ferretduck Sorry for asking, but could that uh, not be in zig? Like, python or something? So it can be used as a comfy node with easy to install dependencies, say. I haven't even heard of that PL till yesterday. Does it perform the same as city's repo - doesn't it need customisation per base?
@ferretduck can you share link to ggufy?
@firemanbrakeneck I'm not sure why the programming language matters, unless you wanted to build it yourself. You can just download one of the pre-compiled releases, you don't need Zig or any of the dependencies installed to use it. I have no interesting in writing Python.
@ferretduck Not so much want to as compelled to once the open source project reaches the abandonware stage, or in order to patch something devs don't prioritise.
Bit baffling, eschewing python in the genai business, but okay.
@firemanbrakeneck Understandable, Python is the GenAI language. I just don't like writing (or reading!) python :) and i wanted to make something very easy to install and use. python you have to make sure you have the right python installed, then install the dependencies, hope they don't conflict with anything or make a virtual environment... Zig you just download the compiler and it takes care of the rest, if you want to build it!
@ferretduck Hello. Unrelated issue, wanted to use the chat but it's broken.
Is it possible to turn ltx-2.3-22b-dev_embeddings_connectors.safetensors into a working gguf?
It seems to be just a massive couple of fully connected layers in bf16, and because it's taking up so much vram I seem to be unable to keep it permanently loaded alongside gemma & main model using multigpu, which should be possible with gguf since I can cap the vram usage. Hopefully. Can't seem to load it directly with the gguf node, many torch errors (invalid argument).
@firemanbrakeneck yes, it should be; ggufy isn't able to do this yet for text encoders but it is a feature I plan on adding.