This is yet another realistic NSFW lora, hence the name. It applies a photograpic realistic look and kinda does it all : AI influencers, everyday fully clothed pictures, erotic pictures, sex scenes, different body types and skin tones.
v1.0:
Yarn v1.0 is here!
This version is based on Qwen 2512, and was trained during 110k steps on a huge dataset. After extensive tests, I'm quite confident this is amongst the best realistic NSFW Lora on Civitai.
Although it was almost exclusively trained on SFW and NSFW of woman-focused photographs, as well as B/G interaction, this Lora is very good at plain SFW realistic image generation.
This iteration is by far the most stable and can be composed with other Loras such as character Loras. All tests were performed using the nunchaku version of Qwen 2512, with the 8 steps Lightning Lora.
v0.3:
This is an intermediate training step of YARN based on Qwen 2512. It performs amazingly well, and is still improving with further training. Pictures often have a distinct "studio look", which I like. I'm even finding myself using the Lora for purely SFW generations.
Pros:
Very good at photorealism
Male and female anatomy are very good
Diverse faces and body types
Composition coherence is improved
Retains high quality when stacked with a character Lora
Cons:
Still issues differentiation between the two holes. This is a captionning issue (I used JoyCaption), which will not improve with further training
Nude men are often missing the shaft if not specifically prompted
v0.2:
I changed the training settings and ran ~x5 training steps. This results in great improvements: v0.2 is much more stable, gives good results at strengths 0.5-1.5 and above. The artifacts present in v0.1 have disappeared. Overall, in my very biased opinion, this is a serious contender for best NSFW Qwen Lora on Civitai.
v0.1:
The lora is a tad too strong at the moment, works best at lora strengths 0.3-0.8.
Known issues : It knows the downstairs anatomy quite well, but can't really differentiate between the two holes in prompts.
Description
FAQ
Comments (2)
It worked with the Nunchaku version (251115) QWEN image edit2509. As you mentioned, it's quite robust, so a setting of around 0.6 is fine. This is the first time I've seen it work properly with the multi-concept Lora. Thank you for your amazing work.
Awesome, glad it worked for you. Do share your results, I'm curious how other people are using the model. I'm working on ironing out the kinks of the model, especially the checkerboard pattern you migh have seen. Stay tuned for a new version in the next weeks probably



















