Pyro's SDXL models - Blowjob edition
Like what you see? Help me create more models by supporting my work. Your contribution can make a real difference. 馃憣
Want more details? Click on 'Show more,' and check out my images for prompt ideas!
Have fun with this model! 馃挦
Hi there!
You probably know me from my SD1.5 models, or you don't, which is fine too...
Anyway, I'm back with an SDXL version of my SD1.5 Blowjob model.
It's pretty alpha since I'm still experimenting with training parameters and such, but for a first try, I think the results are pretty good!
Depending on the prompt every 4-5 images you get something usable. All example pictures are straight out of the model. Did like 30 generations for those pictures.
I, of course, want to improve on this model and am also planning models focusing on non-sex nudity, a POV riding model, and, of course, at the end, a general purpose NSFW model.
But to be frank, this shit's expensive and time-consuming, and therefore your support would be highly appreciated.
Also, if you have a collection of photos you think would make a good dataset that I can use for further training, feel free to reach out
FAQ
How to use?
Just write in natural language what you want to see, like "a woman sucking a cock". No Booru tags yet. Take a look at this rentry to get an idea of what prompts the model was trained on, and use similar expressions. Captions were generated with BLIP, and it somehow likes "big dille" whatever this is.
https://rentry.org/pyros-sdxl-bj
Why a fine-tune and not a LoRa?
So people have a base to merge different kinds of base models, and training the whole model leads to better results. Merge it with one of the nudity models for example as long as I'm not done with my general purpose NSFW model ;)
Where can I find your SD 1.5 stuff?
They were made before Civitai was really a thing so they're hosted on mega:
https://rentry.org/pyros-sd-model
https://rentry.org/pyros-pov-model
Some technical stuff:
Training framework: Kohya's on RunPod A100
Dataset: 3388 images
Batch size: 4
Steps: 22022 / 26 epochs
Learning Rate: 5e-6
Description
Initial release
FAQ
Comments (16)
Thank you!
However the training strength is really high. It would be better to make a lora rather than a full checkpoint.
He can extract a lora from the model. You could theoretically do it yourself too. It is really good though. We need a fully trained porn model, this is genuinely beautiful.
Hi!
Thanks for reaching out!
I'm aware that it's overtrained since that happened consciously, like with my past 1.5 models.
Firstly, it's a feasibility study/proof of concept, not concerned about being space efficient but more concept efficient, and secondly, the model's actual use case is for merging with other base SDXL checkpoints and not general-purpose solo use.
It can do one thing, and one thing exactly, haha! And this one thing it can do pretty good hopefully.
But as the other poster already said, you could make a LoRa out of it - Feel free to do so, and upload it here!
Thanks, Pyro
@pyros_sd_models聽it seems to work well for it's very specific purpose, however, I cannot get anything but bad results when merging this model with other models. I've merged 3 other SDXL NSFW models with each other at varying weights just fine, but merging this model with any of those 3 at pretty much any weight either causes the trigger words to be ignored, or else results in images with a lot of issues, such as bad looking faces, perspective distortions, and bad anatomy, among other things.
As OP said, it seems like this would work better as a LORA.
@seto007聽Use difference merging when merging these specific models, then delete this one to save disk space!
@olivetty聽haven't done distance merging before since I usually just use ComfyUI for merging, though I do have Vlad's WebUI installed. Hate to ask, but what sort of settings would you use to merge this model with another in WebUI?
@seto007
YOURENDMODEL in A
STYLEMODEL in B
SDXL1.0 base model the STYLEMODEL in B was trained on. This will remove itself from B, leaving only the style, which gets merged into A, your desired END model :)
@olivetty聽thank you for the explanation, but using a difference merge still doesn't seem to work out. I tried 9 different merges with the multiplier from 0.1 to 0.9, using both the base SDXL model and a separate NSFW model as C, and none of them resulted in the prompt being included in the image. All that seemed to happen was that the overall image quality and anatomy got slightly worse when compared to the base A model (not as bad as average merge but still).
Given that Loras can be merged into models just as easily as base models can be it still seems like just making this a Lora would be better.
@seto007聽hmm, should work fine. But yeah, a Lora Extract is the best way then and just use the base xl one with the 0.9 vae from official huggingface as the default and it should work fine! :)
Is the "trigger word" really just an example prompt? none of your example images use it...
Hi! Yes, basically. The 'trigger word' is just there for people who can't be bothered to look at the images and example prompts, and who just want to download the models and get going asap.
You did something right with this model. It's insanely good.
yeah, people are sleeping on it for sure.
Thank you for the kind words!
I published a new model! It's not as good (yet), but I would love to see your take on it!
Cheers
https://civitai.com/models/133808
Amazing model but broooo, where is the cum and facials?? It's crazy that noone has made a working SDXL cum lora yet... I know your other model includes some, but this one is much better at BJs
Fantastic model, I am enjoying doing simpleMerge with this and other models.