⚠️⚠️⚠️ 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
FAQ
Comments (37)
<3 this if you think the 3e-6 model is better!
This model is more artistic, the other has more realism in it, even though in this model, the sex positions are weaker as a result.
<3 this if you think the 5e-6 model is better!
I will test both. You are the guy with the gorgeous blowjob model.
@virtualfix6885 Yes I am, and my goal is to have a general purpose NSFW model with the same quality in the future. Not quite there yet, but I'm confident it'll get there with feedback like yours!
So if you have any ideas for improvement, or concepts you want the model to be able to understand in the future let me know!
Thank you!
there's something off with both models. I'm not sure how you trained the BJ model, but that produces significantly better results - even with non BJ images. That said, this does produce some good poses, and decent breast and pussy detail with some effort, but it's spoilt by body glitches and poor detail in other aspects.
i'll play around with it some more after work. Using a character LoRA fixes a lot of the face and body chaos issues I've just noticed.
@lytan Thanks for the feedback!
Actually, both models are trained in much the same way. However, multi-concept training is inherently tougher than single-concept. It's not just about generating just faces and dicks in close up, but different kind of concepts with different needs in sensical details.
The issues you pointed out are exactly what I'm looking to address. And yes, SDXL as a base for nudity is quite tricky since it doesn't really have a built-in understanding of nudity like SD1.5 does as you can try out yourself with the base model. But the challenge to overcome this is part of the fun ;)
That's why I'm currently experementing with different things, and this page is basically my plattform sharing my findings so far
Okay, you should probably add creampies, prone bone, mating press, titjob, deepthroat, reverse suspended congress (full nelson), and cum covered images to the data set, expand the nudity dataset by 20 000 further images, (Just normal nudes without sex, add immense diversity in nude images, AI learns better with diverse images, such as muscular women etc. since it balances itself out) ,etc. 5e-6 appears to have more coherent anatomy than the other model, even though both are pretty good. It might even be a good idea to add rule 34 images into the data set for more artistic outcomes. I'll post some results soon.
i tried merging it with the BJ model and it seems to really add some pop to the eyes and composition. posted some images with NSFW5e+BJ at 65% NSFW. specializing the NSFW models and the merging them into one ultimate NSFW might be the way to go?
trash ... don't waste your time with this :-(
Im all for honest reviews but care to be more constructive?
People spend a shitload of time making these at least we should give them proper feedback
Sorry to hear that you didn't find success with the model.
Can you share the prompts you tried? This way, we can identify and perhaps address the issues you experienced.
Please remember that this model operates with Booru tags, as stated in the model's description and evidenced by my example pictures. You can view a list of tags it's trained on, and hence understands, here:
https://rentry.org/pyros-sdxl-tags
I agree though, that prompting can be enhanced in the future, and it certainly will be. However, as shown by my example pictures, it can produce good results. I had to generate 70 pictures to get those 20 showcase pictures, so the rate of success is pretty decent, even if there's room for improvement.
@pyros_sd_models My only issue with it so far (after playing with it for an hour between work stuff) is the logos and text and urls that it generates on top of the images
@iisjasey969 Thanks for your feedback! I'm aware of this issue, and it will be addressed in an upcoming version. This release was primarily a "quick and dirty" test of SDXL's capabilities without much emphasis on refining the dataset.
In the meantime, using "watermark, text" in the negative prompt should minimize most of these undesired elements.
tons of people "contributing" with a bare minimum effort, doing merges of already existing models around, getting a sh1tload amount of likes. but those who spend hours in training here, those WHO REALLY DESERVE GRATITUDE.. they often get this kind of useless comments and then stays in the dark.
Only god (if exists) knows how much i'm hating this damn monkey brained people comments.
NEVER FORGET THAT IS ONLY BECAUSE OF PEOPLE LIKE PYROS HERE THAT MODELS GET BETTER AND BETTER EVERYDAY, THOSE WHO MADE THE REAL CHANGES. THIS ARE THE MAN WHO DESERVE MEDALS, NOT THIS SH1TTY COMMENTG.
Hey man, if you need a hand with anything shoot me a msg. Not sure if you're captioning by hand but I think that could improve the versatility for sure; and 20000 images is a lot.
Sometimes I am getting unexpected results but its very promising and I like it a lot! Good work!
Tried dozens of stuff from the tag list and even with the "favorite" resolution of 832x1216. All I get is black screens as results although I see something cooking in the preview. VAE setting was ths SDXL_VAE, but even set to auto did not change anything.
Hm, did you try copying the prompts of my example pictures?
I'm also using the official "sdxl.vae.safetensors".
If you use A1111 as a frontend, please try the commandline I use:
set COMMANDLINE_ARGS=--autolaunch --no-half-vae --xformers --medvram --opt-sdp-no-mem-attention
Especially "--no-half-vae" is important.
Awesome! thank you!
one step further for SDXL to shine over 1.5.
Go go go!!
THE BOSS STRIKES AGAIN!!!
we love you Pyro
Pyro is a treasure that keeps on giving. Our most valuable asset in the nsfw wars. Dude is single-handedly giving us the base model of the centuries. Support, support, support this man!
Hi, can I use your model for a merge? I actually had good results improving the faces and the anatomy.
I would give you you credit and link your page.
Keep up the good work!
Just thought to comment, with the 3e-6 version some funny things seem to happen with overall image quality if moving away from the preferred resolution (even at the default 1024x1024). There's an odd distortion, almost like cheap wave filters being subtly applied across details. Mostly evident in faces. It's far less pronounced at 832x1216 but hope it improves with new versions. I wonder if this has anything to do with sampling method as well? I've been using DPM++ 2M Karras with 40 samples. Normally this provides very nice results in sd1.5 and sdxl, curious if anyone else has similar observations.
v013e6 is definitely better for me. It "blends" better with other concepts from sdxl model. v015e6 seems more baked
impressive model hahaha! the Fine-tuning, wasn't kidding hahaha
The watermark for OSOSEDKI.COM is very clearly visible in many of the generations! Maybe apply a blur mask to text in your training dataset?
A lot of watermarks are pouring through. Need a complete revamp.
The body proportions are a bit off in this model, and the lighting isn't great, but I extracted a 256 dim, 3GB LoRA from this. On top of a model with better bodies(but lacking in explicit NSFW), like RealisticFreedom, it reliably generates the best explicit images, with the most variety, that I've seen out of SDXL.
Pyro, you should publish a Lora version of this. If it's easier for you, I can get you the version I extracted.
Can you share your lora please?
I'd be interested in a lora version too
Hey, been following since i found your 1.5 blowjobs model on a "for idiots" style SD tutorial, thanks for your work.
No demo generations i wanna share, but i definately vote for the 15e6 model. ive tested both of them on the own, but my main interest is creating a personal merge involving photography style models and a bit of danbooru friendly. The 15e6 version generally seems to tolerate a wider variety of prompts without generating unhinged nonsense, but i havent tested either version in merges very thoroughly. It could just be my prompting style, but 15e6 seems to be more of a "usable" model.
Thanks for your feedback! Feel free to try out the new version which fixes almost all issues of the old versions.
Dude, definitely extract a Lora from the 15e model bro. IT'S AMAZING.
What base model did you use?
@benjimon678 Juggernaut I believe.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.