Forget my LoRa exists and please use https://civarchive.com/models/1295758/nsfw-fluxorwan-22orqwen-mystic-xxx?modelVersionId=2195978 - as of 9-18-25 - IMO the best NSFW QWEN lora out there.
WARNING- Failed project, training failed, vagina is still not correct. Download at your own risk. Need to go back to the drawing board and try other settings and experiment more. Just leaving this up anyways for reference if anyone wants to attempt their own run.
Pussy/Vagina/Vulva lora training attempt on Qwen with an RTX 4090.
In my testing, setting it to a weight of between 0.75 to 0.9 is good, above that and you may get deformations, and lower than 0.75 and clothes start coming back on to cover the naughty parts. Even though the training run failed Im having a lot of uses with it by mixing it with https://civarchive.com/models/1851673/mcnl-multi-concept-nsfw-lora-qwen-image?modelVersionId=2105899 at a 0.50 weight, by leveraging both nude loras I get better results than it on its own.
TRIGGER WORD- Nude
Trained on only one concept, a nude woman with her private parts visible. I used 24 images to train this lora as the creator of the AI TOOLKIT recommended between 20 and 30 images.
I am not a professional lora maker or AI engineer. Just some random gooner living in his moms basement that doesnt really know what he's doing. So take this lora with a grain of salt.
Ive never really cared much for the harder core NSFW stuff, Ive always been a milk toast 'corn' enjoyer, so a good nudity lora is all I need. QWEN was not trained on NSFW photos at all, even poses that ask for feet up on a fully clothed model gets morphed and deformed, so its gonna be a while before we get anything if at all, so I may as well get the ball rolling. Without widespread community adoption Qwen much like Hidream dies in irrelevancy.
======A few DATASET PROMPTS USED(Used JOYCAPTION)=========
Photograph of a nude woman with light skin and long, straight blonde hair. She is sitting with her legs spread wide apart, exposing her shaved vulva. She has small to medium-sized breasts and a slender physique. Her expression is neutral, and she is looking directly at the camera. The background is a plain white studio setting with a soft, white, furry rug beneath her. The lighting is bright and even, highlighting her body.
A nude woman with long brown hair, lying on her back on a wooden floor, legs raised and spread, feet pointed upwards, hands holding her ankles. She has a slim physique, small breasts, and is wearing a silver necklace. The background includes a black cabinet with drawers, blue upholstered chairs, and a patterned rug. The lighting is warm and soft, highlighting her body and creating shadows. The camera angle is straight on, capturing the entire scene. The image is a high-resolution photograph with a watermark in the bottom right corner.
A high-quality, studio photograph of a nude woman with long, straight blonde hair, standing against a plain grey background. She has fair skin, a slender physique, and small to medium-sized breasts. She is holding a pair of sheer, pink panties with a floral pattern at her hips. The woman has a neutral expression and is looking directly at the camera. The lighting is soft and even, highlighting her body contours. The image is watermarked with in the bottom right corner. The camera angle is straight-on, capturing her full body.
Photograph of a nude woman with long brown hair, fair skin, and a slender physique, on all fours on a white towel. She is outdoors in a grassy field with a wooden log behind her. The woman is looking back at the camera with a sultry expression. The image is taken from a low angle, emphasizing her buttocks and legs. The lighting is natural, with bright sunlight. The watermark is in the bottom right corner.
------------------------------------------------------------------------------------------------------
Watch this video if you'd like to attempt to train too and own a 4090/5090.
In my next run I will try cropping my dataset to just the torso and vagina so the training has less confusion on what it needs to focus on, but im worried doing that will cause the lora to force full body shots to be cropped aka crop out head and feet. In this training run I trained full body shots, which is why it started to converge on a similar face, which is really bad.
Description
MAIN LORA- 3000 steps, likely final version for now. Still not satisfied with the final results, still needed more steps. 3000 looks like it was just finally starting to converge and get good results, but I had set my stop mark at 3000 so my training run ended and now I'll have to do a fresh run, but that's for later.
FAQ
Comments (17)
Just finished training a uncensor+nudify dual concept test lora on Kontext and had the same issues despite training for 17,000 steps. Taking a look at your examples - it seems training this on Qwen would be at least 10 times easier than Kontext if you can do that much with just 1500 steps (as shown in the image). I'm worried about diminishing returns but once (and if) training becomes available for 16Gb - I'll also give it a try. Thank you for sharing this
Use a cloud GPU service like RunPod, super cheap you can rent out GPU as high as 181 GB VRAM (B200) or 48gb LS04 for 0.79 per hour, for 10 dollar you can train up / test and bake a rake of Lora. I was using RTX Pro for weeks before we got the workstation.
What is the method to train qwen? and can it be done on a 3090 (24gb vram) seeing a lot of loras so hopeful but seems to be no guides on youtube anymore for lora training.
yes it can be done on a 24gb vram, scroll down and watch the video in my description box
A bit late reply but saying this in case someone comes across your comment while wondering the same. Qwen can be trained even on a 12Gb VRAM GPU with FP8 quants + block swapping in Musubi Tuner. In the musubi github rep - there is an entire page that covers training for Qwen and what settings one should use depending on how much VRAM one has. Also, block swapping does not decrease training speed by much - I do 5s/it with fully loaded 4Bit Qwen in AI-Toolkit and with Musubi FP8 + block swap I was doing 7s/it (both at 512x512 res on a RTX 5060TI).
it worked for me with a 3090 too and I am just a week into this AI stuff, I have 48gb ram. This one is changing the face of the lady too and giving the image some artifacts.
@javierflores913 This lora no longer being updated and its bad, use mysticXXX lora for qwen
Nice, but you d*n't have to censor the slur
you never know, online censorship is getting worse
Currently training QWEN and WAN 2.2 - with 1924 puntang mages on RTX Pro 6000 workstation (96GB VRAM) - Have acheived God Pussy V1 levels! Coming soon in a front bumper pack, with fine tuned check points....
awesome!!! cant wait to try it out. I cant train with my 4090, its not strong enough to reach the 30k steps and above. When I get to around 3000 steps my iterations per second keeps jumping up exponentially and it would likely take my gpu an entire month to reach 30k steps.
I've just finished a training test on Qwen-Image-Edit for uncensor+unclothing concepts. ~3k pairs dataset (~800 unclothing pairs), trained for ~20k steps (6 epochs with 1 repeat), 512 x 512 res, 3e-5 LR, rank 32 - using Musubi Tuner with FP8 quants on a RTX 5060TI 16Gb.
Compared to the native model, my LoRA can remove clothes while preserving everything else including background, character pose, etc... Its also much better at predicting overall body shape and breast size. But the genitals are still horrible. It still struggles with nipples and its just terrible with puss*. Its so bad I don't feel like its worth sharing.
I'm kinda new at this though - do you have any advice for a noob? Can you share settings and tips for Qwen training? Because clearly training NSFW for this model ain't nearly as easy as something like SDXL or even Flux.
@Jilhbril Woa! Outstanding that sounds epic, not actually had time to try Qwen image edit yet, I heard it could already do the whole nudify thing. What are your training images like? In terms of poonani? are they close up clear images, took a lomg time to break the puntang to get clarity in that "key" area, Now hearing that I strongly suspect that they baked in some kind of chastity belt to block out the "holiest of holies" ...
For us it took a lot of high resolution beef curtain shots to clairfy.
SUPER interested in the Qswen Image edit lora training process now, how did you find so many pairs of on / off images? - I was hoping it would not be like Kontext where you had to train pairs because finding on off photos ruled it out for me.
And what sourcery is Musubi Tuner? Never heard of that one before! Super suprised I thought I knew it all when it cam to training scripts! haha... And did know of any scripts for Qwen Image edit yet, amazing although Its a no go if If i have to find pairs!
@fox23vang226 - Tip on training scripts, you should use AI-Toolkit. Amazing GUI, settings out the wazoo and the best thing about it is you can pause it at any point, check out the epoch images and see how the cook is coming on, if its lacking of overfiiting etc you can adjust any settings then resume, or if training has finished you can start training again from the start using your completed model, so basically esuming and building on top, so if you die at 3000 steps you can just continue from 3000 and go on expenetially, burn it into oblivion, no problem you can check back through test images until you find the epoch where it los the plot and save the model at its last best iteration. Super flexible!
@ronikush is Qwen-image censored by default? it seems so?
Are you setting this Lora or the NSFW one refereced in your notes to 0.50 weight when using them together? It's a bit unclear which one is getting lowered in strength for the best results.
You can avoid my lora all together and use mystic-xxx,
will you update it? maybe with pubic?
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.