⚠️⚠️⚠️ 02/2024 - NEW UPDATE V0.5 ⚠️⚠️⚠️
⚠️⚠️⚠️ NO WATERMARKS - NEW CONCEPTS ⚠️⚠️⚠️
Like what you see? Help me create more models by supporting my work. SDXL fine-tuning is expensive as fuck compared to SD1.5. 👌
Want more info? Click on 'Show more,' and check out my images for prompt ideas!
Have fun with this model! 💦
Don't forget to ⭐⭐⭐⭐⭐ my model if you like it!
Discord channel
https://discord.gg/r2tJpTv4
Join if you want to share prompts and discuss features/concepts for future versions.
Updates v0.5:
(Almost) no watermarks
more concepts
facial
prone bone
vaginal
fisting
feet / footjob (it's a bit iffy)
deepthroat
it's a little bit overtrained. so use a sensible/lower CFG value.
Updates v0.4:
Easier prompting
Better style adherance
more concepts
titty fuck
kissing
some lesbian stuff
cum
more details - sometimes even good hands! wow!
more variety in faces and stuff
Prompting guide
I can recommend to build the prompts this way:
{tags}, {motif}, {style}
For example:
pov, medium breasts, a beautiful woman is getting fucked by a huge penis, professional photo
Use tags for negative prompt to get for example non-nude pictures and so on.
Motifs are “normal language” descriptions of what you want to see like for example:
A woman getting fucked by a huge penis
A fitness model is sucking a big cock
My wife riding a man
See my album for ideas.
Supported tags:
pov - “point of view”
Small breasts, medium breasts, large breasts
Looking at viewer
Penis
Pussy
Ass grab
Kiss, French kiss
lesbian
Nipples
Titty fuck
Cum
Solo, solo focus
Nude
facial
prone bone
vaginal
fisting
feet / footjob (it's a bit iffy)
deepthroat
Prefered resolution is 832x1216 - cfg 5-7
What's next?
Going to try SDXL turbo.
If you want to see a specific concept let me know!
Some technical stuff:
Training framework: Kohya's on RunPod A100
Dataset: 20.000 images
Batch size: 16
Steps: 110.000
Learning Rate: 5e-6
Description
Updates:
Easier prompting
Easier style adherance
more concepts
kissing
tit fuck
some lesbian stuff
more variety in faces in such
FAQ
Comments (72)
New update! Yay! Read description and take a look at the example pictures.
Next version will tackle the watermarks, I swear!
Up until now getting good sexy girls was just more important.
I would love to see cunnilingus done in later iterations... so far no one has done any for sdxl yet, even the nsfw all in one lora doesn't really work on it.
I'm on it! It almost would have made it into this version, but I really lack the amount of training data. Ideally I would need like around 2k pictures for any given concept. So if some kind soul want to donate his collection of images I'll gladly give it a try.
@pyros_sd_models I have anime, otherwise I'd be more than happy to give you what I have. Though even anime I have nowhere near 2k
The new version is INCREDIBLE.
Thank you 😊 If you have any wishes or ideas for a future version, just let me know!
@pyros_sd_models More black,indian girls... basically darker skin women... slightly thicker but not fat.
@KINGLIFER yes, more diversity is on the todo list!
Here are some concepts which would be cool to add.
full nelson (reverse suspended congress)
deepthroat
bukakke (cum covered)
cum shot
prone bone
pile driver
cum squirt
deepthroat and more cum stuff is definitely coming. already preparing the dataset layout for the next training session!
If you have high quality pictures to share, please let me know! I need around 2k pictures for any given concept to turn out well. You can imagine it takes some time collecting those, haha
@pyros_sd_models Just to piggy back on the ideas, would you be open to adding feet focused content for a future version?
Something like this:
https://civitai.com/models/64619/ass-up-wait-feet-or-siren-lora-008
or
https://civitai.com/models/16198/feet-pose-realistic
That might just be the push to get me to finally try SDXL instead of 1.5
If not well that's fine too, I think I'll keep an eye out on this model for when I take the leap over to SDXL anyway, great work! :)
@Gairm Yes of course! It's noted :)
@pyros_sd_models nice work indeed. How can we reach you to help? I have a 1300pics dataset as a matter of fact, oral-themed. Would be better inside your model than one more random LoRA. I need some instructions as to what you need in terms of formatting etc
@pyros_sd_models alright will do that later today, thanks!
@pyros_sd_models i second the request specifically for prone bone. haven't found anything on civitai that can do it yet, and it's my fav position IRL :)
@pyros_sd_models Do you just want pictures or should I add a text prompt as well? Like "woman having full nelson sex" and "woman fucking full nelson style" or should it have booru tags like "1girl, 1boy, full nelson" etc?
@pyros_sd_models Do you just want pictures or should I add a text prompt as well? Like "woman having full nelson sex" and "woman fucking full nelson style" or should it have booru tags like "1girl, 1boy, full nelson" etc?
@virtualfix6885 It would be more than enough for me If you would organize images in different folders. For example all bukakke pictures in a bukakke folder, all prone bone in a prone bone folder and so on. That's enough for my scripts to do their magic on :D
If you have more questions pls msg me pyros.sd.models (at) gmail.com
I guess it would be a good idea to start a discord, haha.
I don't know if its an artstyle you're going for, but there's just this strange look to your models, the faces and proportions, everything is just a bit surreal, like something a SD1.5 model would put out, except this has XL style sharpness and can output cocks better. Just simple one word prompts like "girl" or "woman", and you can see this model is built on top of something very strange. Maybe its because I'm on A1111, and comfyui outputs something completely different.
I mean just writing "girl" or "woman" into base SDXL will also give you pretty surreal stuff, haha :D
The only thing I can think of is that there's a distinct look with the girls it produces because plenty of the dataset consists of - how the zoomers would call it - "thots". You know all those "too perfect" onlyfans and instagram girls for which you could argue that they all have kind of a plastic-y surreal look.
But that's by design, because I think they're hot.
My images are also all made with A1111
@pyros_sd_models I assume it because you haven't tagged them with natural language prompts. For me, writing "girl" or "woman" into a model is my first test of the training style. Those that were trained cleanly, like Afroman or Freekhitman's models can output something pretty decent just off of that. My own merged models included.
Still, great work. Glad to see you pushing the boundaries of the models context once again!
Best NSFW SDXL model so far, mainly because styles work amazingly well contrary to other recent NSFW models. Please, whatever new NSFW concept you add, don't break styling capability :)
Thank you! I will try my best :)
Please feel free to also give me some review stars which helps with visbility on civitai!
Discord channel
https://discord.gg/QGnSYWsr
Join if you want to share prompts and discuss features/concepts for future versions.
link doesn't work. need to set it to permanent in discord
one word, Amazing!
How are you guys getting cunnilingus to work? i cant get anything for it. Even copying the prompts from here i cant get it to work.
I don't know either. It's not a concept the model is trained for yet, there's literally no cunnilingus picture in its dataset. You could try "a woman licking a pussy" or something. The model knows what licking is, and perhaps you can get it this way.
Awesome model and my instant favorite!
What I can't find is any "backdoor action", or I'm not using the right tags. Is there a Lora for SDXL which could do it?
What exactly do you mean with "backdoor"? anal?
The model is not explicitely trained to understand "anal" since in my early testing I noticed the model has a hard time to differentiate between anal and pussy sex. It's too similar so there's much "concept bleeding". so currently it's just "luck" if you get an anal picture.
Perhaps I give it a try again in a future version!
Try like (detailed anal sex:1.5) and things like penis sliding into asshole
@pyros_sd_models yes please
Currently the best SDXL model!
I agree
how tf do you fine tune SDXL my 4090 is struggling to do it
Loras are perfectly trainable on my 4090, but I didn't try full finetuning. But they didn't use a 4090. It's on the description: They used runpod with nVidia A100 80GB. It's 1.99$/hr according to their pricing table, and something like this probably took a lot of time to train... So not exactly cheap, I guess. 😅
Hi there! As the other person pointed out already I rented GPUs on runpod and used kohya's training framework as a base for my works.
I'm getting watermarks on 99% of generated images, even with negatives. Edit: Saw your comment about future updates.
In the mean time you can just go to inpaint and inpaint it out or use photoshop.
Could you please update the Discord link? Says "The invite is invalid or has expired" to me. Thank you.
Link is updated!
Please add thiz to mage.space
This is a sort of a crazy model. Very CRAZY Results, but I somehow like it.
However, want it more stable =))
please do a LCM version! love your work, just phenomenal!
Difficult to use, but insanely good
I'm having a lot of trouble getting the penises to appear fully inserted, whether for a blowjob or for sexual acts. Any suggestions for this?
Most likely the issue is that you have the word "penis" or "cock" in the prompt, and that makes the sampler want to draw you a penis or cock. Which means it can't be inside whatever you want it to be inside. Your prompt needs to "talk" around the penetration without using the actual word of what you want to be disappeared inside a mouth or vagina. Utilizing a more natural sentence-based prompt rather than the old school tag-style will help a lot with this. Remember that SDXL is an enhanced language parser. It can read. "a pov photograph from above of a woman having an orgasm from deep penetration during sexual intercourse" rather than "missionary sex, cock in pussy, orgasm, deep penetration,"
It's my fault somewhat. at least for blowjobs the training data set doesn't have that many fully inserted blowjob images. Next version will know what deep throat is :)
also what @shapeshifter83 said.
For example this prompt doesn't have the word "penis" in it, and in the resulting image it looks like it's fully inserted.
https://civitai.com/images/4332623
@pyros_sd_models during your training, consider the phrase "balls deep" for total penetration of vagina/anus because that's a somewhat "common" utterance that also isn't likely to confuse the sampler word-for-word, ex. "balls" and "deep" won't throw a sampler off like "cock in pussy" would. Also, for this same reason, consider "deepthroat" token rather than "deep throat" tokens because with a lot of typical pov images, you don't want the sampler trying to draw a "throat", as that would usually be hidden. Just my 2 cents!
Thanks for the replies! I haven't had much luck using more natural prompts. For instance: requesting the perspective from under a table puts the subject under the table rather than the "camera". I've tried rewording it a few different ways but still get the same output. Also good work Pyro, looking forward to the next checkpoint :)
@shapeshifter83 Thanks! I'm aware of the text encoder intricacies, and of course for the upcoming version "deepthroat" will be the key word and not "deep throat" for the reasons you stated :)
with "balls deep" there's the problem that the algorithms I use to caption my training set just doesn't recognize the difference between fully inserted and not, so I would have to manually sort out 30k images, Fuck no, haha :D
If you peeps have already a collection of pre-sorted "balls deep" images I of course would love to get it and integrate it into the model :)
@bizarrenarnar99640 Yes, but that's generally a StableDiffusion problem. It's not smart enough for complicated compositional tasks. For example the base models already has troubles to do something like "a red ball on a blue cube".
"from above" and "from below" are two expressions you can play with to manipulate the camera viewpoint
I cannot get consistent good results from this. I tried to follow the prompt syntax, but the the output is always horrific. Indescribable body parts growing out of different places. Disfigured hands and penises. Every now and then a good render emerges, but it takes a lot of tries. Even when I copy prompts posted here, it's inconsistent. Really difficult to use IMHO.
The next model should be out soon. So keep an eye out and follow Pyro on discord.
I got exactly the same issue.
Love the model and can't wait for the next version! Discord link seems to be invalid though, can you update the link?
Updated: Jan 4, 2024
Have you uploaded the new version, the early access tabs are all gone, thought I'd be able to download the new version soon.
There is no new version coming anymore, I think. Pyro abandoned the project, I think.
We must find solace in the last remnants of what remains of our blessed compatriot. :,(
@virtualfix6885 What? NOoooooo
He's also put so much time into his new model that it's all trained up, even if it won't be updated again. It would be nice to see the results of this one and everyone would be happy. After all, sdxl is the only one doing nsfw fine tuning.
@suede2031691 there are some of the new pics in his pictures. I sending some of my dataset + watermark remover to other model creators.
Also, AnimagenXL 3.0 is really good at nsfw etc. You can look at my pic history and check how good it is. It's like when NovelAI got leaked or the other anime model that I forgot, was released.
@virtualfix6885 Aww too bad. This was/is my fav NSFW model. Is there a reason why? Is there a new project? Is there a discord or something? Sorry for all the questions.
@foveh76999518 I am in the discord, https://discord.gg/jZRPaDAU
I cri evertim :,(
Nice working model. I dont have any unsolvable problems to use. Working good with anothers LORAs. The preliminary result of the very good level of generation
99% of photos contain a link to the website, text, logo etc...
Try the new version :)
@pyros_sd_models thank you, v0.5 is much better :D
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.