I trained a few LORA's using hundreds of images from a popular nsfw subreddit to produce realistic female butts and vaginas/labia/genitals.
I'd recommend downloading them all and trying them all out.
(I see ya'll only downloading the first LORA. This is fine, but you might not get the desired results using just one)
I manually captioned every photo and used some WD1.4 autotagging as well. I have no idea if I captioned the images properly. I have no idea wtf I'm doing, heh.
Side note: Manually captioning images absolutely sucks. BLIP and CLIP currently do terrible jobs at captioning NSFW images.
These LORAs were trained using https://civarchive.com/models/2661/uber-realistic-porn-merge-urpm (1.2) as a base model. You should probably use this model with these LORAs.
I also found it sort of works with HassanBlend 1.5.1.2 (after forgetting to switch models). It may work with other models made for photorealism, but it might also produce some horrors.
You should also use the bad_prompt_version2 TI (Textual Inversion) with this as well:
https://huggingface.co/datasets/Nerfgun3/bad_prompt/tree/main
Turns out when I was testing these LORAs, I didn't have the bad prompt TI installed (I reinstalled SD and forgot to re-add them). I'm not sure what "bad_prompt_version2" was actually doing in my negative prompts...
None of these LORAs are perfect and it might take a bit of text2img and then img2img and tweaking the prompts to get good results.
You can use them individually, paired, or all three.
Setting the LORA weight to 1 may fry the results. You'll get better results keeping them around 0.8 and maybe lower.
Set your GFC anywhere between 3 to 5. Anything higher will probably deep fry your gens and may produce some...interesting results.
Genning with PLMS @ <40 steps seems to produce better results. Using DDIM and DPM Adaptive seem to work pretty well in img2img and can sometimes help add realistic skin details back.
This model is not very good at faces from the side, even with Restore Faces checked (this might make them worse).
Hands are sometimes OK.
Honestly, just play around with it, or merge it with another model or other LORAs, or whatever.
Just don't use it for evil, to intentionally trick, deceive, and/or manipulate people for personal or group gain, illegal whatever, etc.
I have literally no idea wtf I'm doing with any of this. I am not an expert or even close to one on this subject so please don't ask for advice how to make your own LORAs or whatever.
Neither myself nor Civitai, or any other model authors/mixers are responsible for any damages to the living, deceased, or to any property caused from using these LORAs, in part, or in whole. You are solely responsible for what you do with these models and the consequences thereof.
Description
Trained on an assortment of 169 pics
FAQ
Comments (12)
Hi! Looks like a great model. There's a similar one called subreddit, would you think this one does a better job? also, is it better for inpainting or does a good job with txt2img with embeddings for persons?
I was unaware of the Subreddit model, but it looks to be a lot more diverse in terms of what it's able to generate since it seems to be trained on a bunch of different NSFW subs vs. the 3ish gonewild ones my LORA's are.
I'd def recommend giving it a shot as well. I know I will be.
I haven't tested my LORAs for inpainting or embeddings with people, but I'd imagine they'd be the same as the base Uber Realistic Porn Merge (URPM). I would guess these LORA's should work OK on the URPM 1.2-inpainting version, but I haven't tested it.
@MisterMan117 how do you make this ? i have 84 GB on NSFW content. dont ask me why/how lol i wanna create the most realistic masterpiece lora/model
@chrissy22 First you need to do a selection, try to separate simple nudes from porn, then you have to separate what type of porn: doggystyle, blowjob, etc... Then you have to run a caption thing on them... I guess correcting wrongly put labels would be hard, so I'd probably skip this part... then run lora training on them
It looks to me like you've trained this on a number of low res images. Am I correct? I say this as low res images seems to be what this model can put out. It absolutely does what it says it can do, but it makes me wonder how the model would have done if trained with high res images.
just add "lowres blurry image with compressed jpeg artifacts" to negative prompt
I tried training another LORA with 768*768px images and the results were about the same as this + upscale. I'm betting I'm doing something wrong. I literally have no idea what I'm doing...
i only get anime results, any ideas?
what are you using as your base model?
What is the trigger word?!
Thank you for your work. I'm sure most who benefit will not contribute back but I've been impressed.
How may i use the prompt for this Lora? Kinda new and want to incorporate this in a prompt. Is there a way universally to include Lora prompts that I may not know of?