This is an experimental LoRA that does one thing relatively well: PIV sex. You have to describe the scene precisely (see the prompts for example images) and sometimes you need to remind it that everyone is naked. It doesn't really know the names of sex positions (if you say cowgirl, you get a girl in a cowboy hat, for instance).
That being said, it's a LoRA that trains Qwen Image, which is otherwise completely ignorant of sex, to do pictures of sex.
It can also do female nudity, but despite the presence of dick pics in the training data, it doesn't seem to understand dicks outside the context of being in vaginas. I think a full finetune will probably be required to do any better than this, so I wouldn't necessarily expect updates.
Description
FAQ
Comments (20)
Can you write the prompt words for behaviors or actions recognized by lora in your post? I don't see the prompt word for your picture. Thanks.
none of the examples has the prompt in them just the lora its using
Argh, so frustrating.
If you want to see the exact prompt, save the image and drag it into comfyui.
Otherwise, here's the Dynamic Prompts template that I used:
1girl, {naked|wearing an unbuttoned, open shirt|wearing a bra and no panties|wearing a miniskirt that is lifted up past her waist exposing her vagina|nude|wearing lingerie}, with hair, full body. The woman is having sex with a completely naked man, who is {standing in front of her squeezing her butt with his hands while she stands on her tiptoes and grinds her crotch into his|lying on a bed and she is straddling him and leaning back {facing him|facing away from him}, arching her back|sitting on a couch with her in his lap {facing him|facing away from him|facing the side}|sitting in a chair with her in his lap|standing behind her while she stands on her tiptoes and bends forward at the waist, arching her back|{standing on the ground|sitting|lying} with his crotch between her wide open legs and his hands on her {butt|hips} as she {lies|sits} on a {bed|table|couch|counter}}. {She is blushing and her eyes are closed.|She is blushing. Her mouth is open and she is screaming out in sexual pleasure and her head is tilted back.|She is blushing and having an orgasm.} His penis is inside her vagina. Both of her legs are visible and spread apart.
his clothes: completely naked
Obviously you need the Dynamic Prompts extension for the template to work, but even if it's not installed, you can get a good idea of how I prompted it.
This was all going on during the confusion thinking that Diffusers or something must be pulling a Reactor Face Swap on me, only to find out that Qwen is allergic to Sage Attention and throws a black screen if you have brackets.
I'm not sure about the benefit of the LoRA, but I noticed two things in your workflows:
1. <> used in prompting, allowing negative prompting even at cfg 1. There's also underscores and shit I'm going to figure out sometime
2. I originally found this looking for ways to draw tall and thin characters, and I noticed that some of yours mentioned "disney style cartoon" in the randomized prompt results shown in the lower left. I tried "Disney character proportions," and now I can have tiny waists and six feet tall supermodels without a single LoRA!
Thanks for the ideas!
It also makes their hair long enough!
Other terms that seem to work:
bruce hartman character design
stephen silver character design
classic disney villainess style
disney renaissance character art
fashion illustration style
jessica rabbit pin-up style
These prompts work on Wan, too!
You may want to train the next version with more images from behind.
It worked well but it has a problem with the toes bleeding into the back sole of a barefoot character.
And the Pussy from behind looking like the there's a fleshy bit hanging from it.
Huh, I haven't seen that. Try using more clinical langauage, like vagina. Pussy can also mean a cat, and that can confuse Qwen, the same way cowgirl and cameltoe do.
@_Envy_ No, I used the correct terminology in the prompts. And after more trys It does look like it messes up the toes when asked for barefoot characters, far more than base Qwen.
Could you share some info on how you trained it?
- How many images in the dataset?
- How did you caption the images?
- What tool did you use (AI-Toolkit/Musubi Tuner/etc) ?
- What training settings did you use?
- Also if you were to train for the first time with all of the knowledge you have from your experience - would you have done something differently?
- 500-600, give or take, a mix of manually curated anime generations from various Illustrious checkpoints and screenshots from pornhub videos
- Mostly with joycaption, some left blank, some captioned manually
- AI-Toolkit, with these settings: https://pastebin.com/RqGyhv9y
- Probably double the number of dims of the lora, because hopefully that would allow it to understand more concepts.
@_Envy_ A bit confused about your learning rate value of '1'. If it really trained on that then its super hella over-fitted even after 500 steps let alone 60,000. Maybe AI-Toolkit defaults back to a lower value when something like that is given? What is the reasoning behind it - if you don't mind me asking?
Also thank you for the reply. It does give me some insights.
@huj0ps1t6 Its prodigy as optimizer. Thats why its 1.
@huj0ps1t6 Prodigy is supposed to handle the hyperparameters for you. That’s why the LR should be 1. As for the schedule I had great success using constant or cosine. If you have a learning rate and schedule using Adam that works for you, there is no need to use prodigy. In fact I believe it uses more VRAM. That said, prodigy might converge faster and you might need to check earlier checkpoints by altering your save-after-steps.
@LDWorksDV TIL. Thank you. I did notice it was using prodigy and that's something I never actually tried. I've been using adamw8bit with Musubi Tuner and lr 2~5e5 with reasonable success in Qwen-Image. It may require more steps but for a dataset with several hundred/few thousands of images I find that value quite reasonable. So far the only struggles I've encountered are NSFW related hence why I asked OP for some insights.
@huj0ps1t6 It's not lobotocensored like Flux is, but it doesn't seem to have been trained on much NSFW stuff at all, so you're teaching it entirely new concepts (the text encoder seems to understand literal, clinical terminology though).
Can you release the training dataset?
My question to this is ... why?
It doesn't really do anything that even an SDXL-based model can't do. Why wait longer for QWEN to do it (with errors, even, because it doesn't natively support NSFW)?
because qwen can maintain much more complicated prompts.