NSFW Lora for Qwen Image 2.1 (V2)
In my opinion this is currently the best NSFW Lora out there for Qwen Image 2.1!
What i recommend for the best results:
25+ steps
3-6 cfg
0.8-1 for lora strength
er_sde for sampler (if i use euler i get very bad results!)
beta scheduler
Try to test different types of prompts to see what works... Some work really good while others don't... I'm trying to fix it in a v3 but it takes some time! :)
Lora (v2) is mostly trained on: (these are not triggers)
Blowjobs
Cowgirl
Cum
Vaginas
Anal
Penis
Doggystyle
Missionary
Dildo
Pussy Licking
The qwen image 2.1 model its license doesn't allow earning money directly...
So if you want to support me:
Tip me some buzz.
Tip via https://ko-fi.com/thesealpacas
(This is not at all required, but this really helps me make more free loras faster and better!)
I made my own discord server where i post tips, updates and extra models!
Please let me know what works and what i should try to improve! :D
Description
Experimental version.
FAQ
Comments (61)
I was having trouble creating NSFW character sheets and this helped. Thanks!
Great to hear! Tysm!
Works so good!! Thanks for adding this.
Seems to be only trained on one age and body type, pushes every prompt toward that, overrides face and age prompting before it effects nsfw prompting. Useful for that one type of character, hard to use it for anything else.
Hmm... i will look if i can fix that!
Hey, i looked into it but i don't rlly understand what u mean... I can generate both fat, skinny, young, old, skin tone, ethnicity. without issues... Did u use er_sde and beta?
@TheseAlpacas I use prompts with very specific language to describe macro facial features for repeatable characters without lora or ref. This lora pushes the faces away from the descriptions, generally toward a much more youthful instagram face.
@stat83 ah i see... well i dont rlly know how to make it so it keeps that in tact... U could try to generate a image without lora of the character and then use it as a reference image to generate with lora?
works decent for standard image gen but for edit it completely shifts the face
yea just getting body horror on the edit mode even with suggested scheudler and sampler
Hmmm... i will look why that happens...
As far as I've tested for NSFW edit only video models work.
idk edit image turns out fine maybe some morphing
@TheseAlpacas
I'd take another look at the dataset first. If you have a lot of close-ups/faces, mostly the same person or similar-looking people, body horror/bad anatomy images, etc., the LoRA can start picking that stuff up pretty strongly.
Captioning can also mess with it quite a bit. Stuff like captioning someone as kneeling when they're actually standing, not captioning important things, inconsistent/messy captions, or accidentally describing the same thing twice.
Like J1B mentioned, the LoRA also just looks pretty strong in general. Could be overfitting from not enough diversity in the dataset, too high LR, too many epochs/steps, etc. I'd play around with those, and maybe optimizer/rank as well.
I'd also definitely try cropping and/or masked training. Crop faces/heads out when they're not relevant to what you're actually trying to teach and caption the crop accordingly. For images where you can't really crop without losing important stuff, mask out the parts you don't want it learning (with masked training). You could also just try masked training in general and see if that helps.
@Kenzato I will try some things out! Tysm for the tips!
I'm really struggling to write prompts. Do you have any suggestions?
type in what you want on an llm and have it improve the prompt for image generation
Yes like pogblaids3876 said, try to use an llm to expand ur prompts. for me even qwen3.5 4b hauhau uncensored on q4 works fine!
@TheseAlpacas @pogblaids387 I'm doing this, but I don't think I have a good enough system prompt.
Good Job！and thank you~
Sorry, I got a bit carried away with comparison posts for this lora (I have about 100 more not postred as well!) , i was just excited to test out the new model, everyone gets it now, pussy's still need some work and the lora hurts detail/realism a little bit, but it is definitely a step in the right direction, I was mainly using it between 0.65-0.8 as it is a little strong for my liking. butt plugs need a lot of work or a dedicated lora. Good job
Tysm for your comment! This model was trained to have penis visible, as i found that was the biggest issue with base model. The vaginas where "ok" in the base model so i didnt trained too much on it. And i didnt include butt plugs, dildos or anything like that too... So that doesnt work as good too sadly! I have trained a higher resolution lora of this but i didnt fully tested it yet so im still working on improving this lora! 😁
anyone knows how to connect the lora node? its my first time using qwen 2.1
U use comfyui?
If so, just add a load lora node. And link ur model loader to the lora input. And the lora output to the ksampler.
@TheseAlpacas thank you very much!! :)
@nenecaliente69 ofc! 😀
These Alpacas Wow 🤯
Not sure what am I doing wrong but this lora gives 0 changes when enabled VS no lora at all. On Reference to Image task
Huh thats weird... But it does work on text to image or?
check your terminal if the “LoRA key not loaded” warning exist when loading the lora and generating, if it there then the lora is ignored hence the 0 changes.
There is indeed a 'LoRA key not loaded' warning. What is the reason for this?
@aoshanrongyu1133 im checking why
@aoshanrongyu1133 i dont have the "LoRA key not loaded" issue... mind sharing me ur workflow?
@aoshanrongyu1133 try use the native comfyui node "SamplerCustomAdvanced" rather than ksampler. In my case the lora is loaded and the warning is gone when using the SamplerCustomAdvanced node.
@dolores332 for me k sampler works fine tho... so idk what causes the issue.
I adjusted the weight of the accelerated Lora to 1 or turned off the accelerated Lora, and now no error messages appear. You can try to see if it solves the problem.
@aoshanrongyu1133 u use another lora on top of it? maybe thats why... i dont know how to fix that
I uploaded a picture that contains the workflow I am currently using. I hope the author can point out the mistakes in it, thank you very much.
Ah yes i got:
[WARNING] lora key not loaded: diffusion_model.blocks.0.cross_attn.k.lora_A.weight
[WARNING] lora key not loaded: diffusion_model.blocks.0.cross_attn.k.lora_B.weight
[WARNING] lora key not loaded: diffusion_model.blocks.0.cross_attn.o.lora_A.weight
[WARNING] lora key not loaded: diffusion_model.blocks.0.cross_attn.o.lora_B.weight
[WARNING] lora key not loaded: diffusion_model.blocks.0.cross_attn.q.lora_A.weight
[WARNING] lora key not loaded: diffusion_model.blocks.0.cross_attn.q.lora_B.weight
[WARNING] lora key not loaded: diffusion_model.blocks.0.cross_attn.v.lora_A.weight
[WARNING] lora key not loaded: diffusion_model.blocks.0.cross_attn.v.lora_B.weight
and a lot of this in console. And as i said the results is pixel identical with using lora vs not using it.
I am using regular KSampler with Load Diffusion Model -> Load Lora -> Patch Sage Attention KJ -> Qwen Image 2.1 Cache -> KSampler . Nothing special and no additional loras.
base model is:
qwen_image_2.1_int8_convrot.safetensors
@aoshanrongyu1133 it seems to be when 2 loras are used. like a turbo lora together w this lora.
@vladmight875 maybe the sage attention? let me check
Just tried V2 it loads correctly and does some difference, funny enough it changes perspective, camera angle, and a bit of body structure of the reference(woman) but not the genitals leaving them as ugly and non realistic as it were rendered without lora.
i am getting the lora keys error too, no change when alone or not.
@lhucklen what workflow u use?
can you share the workflow for both image generation and img to img edit please.
Sure! Gimme a sec!
I dont really know how to add the workflow. But its just the comfyui template with the lora node added... U can drag and drop one of my images and most of them should have their workflow attached.
Why does it always result in anal sex... even when the prompt specifies missionary or vaginal sex?
🌹
Hmm... i will look into that! Working on a better version!
nice work! thank you
i m still getting broken pussy, what is the parameter that i should use?
I noticed that using a higher cfg seems to work better! For the rest just follow my recommendations! If u need more help lmk!
@TheseAlpacas that really nice~ millions thanks
How do I serve it with vLLM omni? I have tried with many parameters
Uhh... i never used it! I can look into it later...
Nice V1. Just wondering if your dataset is diverse? Do you only train on nude full body transformations? Or does your dataset include, portrait, medium shot, long shot, clothed to full body transformation? If not maybe you could add cropped images from your dataset and/or use qwen 2.1, Flux.2 Klein or any other model to put some clothes on some of the source images? Maybe you are teaching the model very narrow transformation skills? One last thing, your LoRa seems to lack detailed knowledge about female anatomy, seemingly lack of diversity, what it can do, close-up, medium and long shots.
It should have different shots and clothing. And yes i didnt trained this on female anatomy. As i found female anatomy was "ok" on base model. Im working on v2 fixing most of these issues tho!
@TheseAlpacas Also you should make sure to adhere to the official Qwen Image 2.1 edit prompt instructions "https://github.com/QwenLM/Qwen-Image-2.1/blob/main/prompt_rewrite/prompts/system_prompt_edit.txt".