NSFW Adv. is a LoKr trained on Qwen-image FP8, focused on refining and adding details to the human body. It was trained on a curated dataset of over 3000 images. I had to scrap it and start over from scratch more than 10 times, with this final version being trained for over 100,000 steps.
Even with all that effort, it's still got plenty of flaws. But after grinding on it for over two months, I had to push it out. The whole process was honestly super frustrating—tweaking, testing, culling, and repeating the process with a 10,000+ image dataset... it was just so damn hard. I seriously thought I might never finish it. But anyway, it's finally out. Hope you guys have fun with it.
IMPORTANT USAGE TIPS:
First, you'll need to install this plugin for your ComfyUI: https://github.com/ClownsharkBatwing/RES4LYF
Once it's installed, you'll get a bunch of new, high-quality samplers and schedulers in your ComfyUI.
I recommend using Sampler: res_2s and Scheduler: bong_tangent.
I also recommend using it with this distilled LoRA for faster generation:
https://huggingface.co/lightx2v/Qwen-Image-Lightning/resolve/main/Qwen-Image-Lightning-8steps-V2.0-bf16.safetensors?download=true
All demo images were generated using:
Steps: 10
CFG: 1
Sampler: res_2s
Scheduler: bong_tangent
NSFW Adv lora strength 1.0
Qwen-Image-Lightning-8steps lora strength 0.4~0.5
PS: If you plan on merging this, please give credit or link back to this page. Thanks
PS2: This LoRA training session was done with https://github.com/ostris/ai-toolkit, but I incorporated the new sampling methods for timesteps from sdbds (https://github.com/kohya-ss/musubi-tuner/pull/407).
After several experiments, I found this method to be significantly effective for this training task, as it notably accelerated convergence. A big shoutout to them for their excellent contribution.
Description
FAQ
Comments (40)
您工作流中的"OpenAIChat VisionAPI"是私人節點嗎? 找不到來源出處。
Yeah, that's a custom node I made. It's nothing too fancy, just hooks up to a third-party LLM API to enhance or rewrite the prompt. I do it mainly to avoid accidentally using captions from the training set, so I can properly test the LoRA's actual performance.
@fok3827 "Hi, I am using your ComfyUI workflow. However, like the previous commenter mentioned, I am running into issues with the private nodes.
Could you please provide a detailed step-by-step guide on how to connect OpenAI or Gemini? I really want to get this workflow running properly. I would appreciate your help. Thank you!"
@fok3827 my E-mail:[email protected]
@yabo9263 Fok_OpenAIChatVisionAPI is just a custom node I built to connect to an LLM for prompt enhancement—since the affordable API I’m using doesn’t have a prebuilt connection node. You can swap it out with any similar node; it’s not a critical or irreplaceable component.
You mention flaws, but this is maybe the most successful NSFW lora I've seen for Qwen Image. Great work.
Awesome lora! Interesting that you recommend the lightning lora applied at a lower strength, I tested with 0.5 vs. my usual 1.0 strength and it doesn't seem to make a difference?
Using a higher strength for the 'lighting' LoRA can give you more accurate limbs, but it tends to destroy the details and the overall image quality might suffer. For a more natural-looking result, I recommend not setting the strength too high.
Are you using RES4LYF as recommended?
@ruGlory yes
G R E A T ! ! !
The most authentic pubic hair. Could you make a male dick lora?
你满不满意我不清楚至少我觉得很不错了很棒，可以分享下你的通配符吗这是我的邮箱不慎感激[email protected]
太强了,你的nsfw是c站唯一适合亚洲人的
i'm not exaggerating, the results of this shocked me! your hard work was not to waste. works fantastic with edit 2509 also
how? with qwen edit, don't work for me
work only with t2v?
This is a Qwen Lora, it's for images
Usually (always?) regular QWen loras work for 2509 as well, though there is special training for 2509, so not sure what the difference for LoRA specially trained for 2509.
wow,this is perfect qwen lora.
i honestly already downloaded literally all qwen lora for this type and this works wonderfully.
Asians...
Did you integrate logsnr or qinglong? And would it be possible to share your modifications?
Yep, I ported the Qinglong algorithm to ai-toolkit and made some experimental changes. The whole thing was coded by Claude.
great lora lots of nsfw concept, but seems not really anus stuffs
Recently found that a strength of around 0.4 significantly improves face consistency
This LoRa doesn't work when I test it in the NunchakuQwenImageLoraLoader. How can I fix this issue?
maybe try using regular lora loader?
Nice
Its like a bit blurry and smudgy for me, like it looks like a image taken on cheap camera or something, is it supposed to be like this?
Same, actually getting better results using Euler as the sampler and "simple" as the scheduler. Using bong tangent I get really blotchy skin and lots of noise.
@shicken684733 well I am using the z-image-turbo now the quality on it is rly rly good and its decently well uncensored but doesn't understands concepts of nsfw so just waiting on lora's, not to mention its really really fast like 50-70sec on 4070 with 2048x2048 (1024x1024 with a 2x 2nd highresfix pass)
Does this works with qwen edit! Or is it just for t2i
我觉得不适合
步数：10
CFG：1
采样器：res_2s
调度器：bong_tangent
NSFW 高级 lora 强度 1.0
Qwen-Image-Lightning-8steps lora 强度 0.4~0.5
为什么完全按照设置像有马赛克
用的是qwen image还是image edit？
Thanks for great job, your lora is fantastic! Especially thanks for RES4LYF recommendations, it works great!
Does this LoRa have a Z Image version?
Anyone got this to work with Wan2gp?
No trigger words ? (this might be a stupid question, i'm very new to this)
The god knows that when importing images, there is a workflow issue: “Fok_OpenAIChatVisionAPI”，What node is this？
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.