Noct Q
Qwen Image 2.1, uncensored out of the box, packaged for ComfyUI in two versions: Turbo (6 steps) and Base (full steps). Nudes and explicit scenes render without a LoRA.
What's new in V3
Turbo rebuilt and sampled in 6 steps: sharper images and seeds as varied as the full model
In a blind test on the same 150 prompts and seeds as V2 Turbo, fewer defects with the same prompt adherence and explicit rendering
Base is a new model, stronger on explicit scenes
Settings V3
Turbo: 6 steps, euler, no CFG, 832x1248.
Base: 25–50 steps, euler, simple. V3 Base: cfg 3 with the negative prompt, both set in the workflow.
Prompt examples
The prompts of three V3 Turbo showcase images, written like captions of the finished picture: the medium first, then the woman, what she does, where, the light and the framing. All three use V3.0 Turbo int8 convrot in the V3 workflow (6 steps) at 832x1248, and the same seed should give the same picture.
Attic room, seed 96060
An amateur photograph of a nude woman in her twenties with long light-brown hair and full hips, kneeling on an unmade bed in a Parisian attic room with two round windows, stretching her arms above her head with her eyes closed. Morning light through the round windows, rooftops outside. Full body from the side, sharp focus, natural skin texture.
Candlelit bath, seed 96120
A phone photograph of a nude woman in her twenties with a blonde bob and a slim build, sitting on the edge of a deep stone bathtub surrounded by lit candles, testing the water with one hand and looking down at it. Warm candlelight, dark slate tiles. Full body from the side, sharp focus, natural skin texture.
Oil painting, seed 96240
An oil painting on canvas of a woman in her twenties with auburn hair pinned up loosely, wearing a long white linen dress, standing by a tall open window in a quiet Dutch interior and reading a folded letter, her face turned down toward it. Soft north light falls across her face and the dress, a blue-and-white jug stands on a wooden table and the floor has black and white tiles. Visible brush strokes on her skin, hair and dress, a muted ochre and blue palette. Full body
What it does well
Full-body nudes in daylight, flash and night scenes
Explicit scenes, oral most reliably
Readable text on signs and labels
Anime and painted illustration
Plain-caption prompts: who, what she does, where, the light
Speed: about 12 s for a 1024x1536 image on an 8 GB card with int8 convrot
Image editing with the same files, in ComfyUI's Qwen Image 2.1 edit template
Which file by Vram
8–12 GB: int8 convrot (7.3 GB), the default, almost pixel-identical to bf16
6–8 GB: int4 mixed (4.6 GB), the fastest; compositions shift a little
GGUF Q8_0, Q6_K, Q5_K, Q4_K (7.6 to 4.1 GB), in v1.0
fp8 (7.3 GB): plain fp8 e4m3fn
bf16 and fp16 (14.2 GB): for merging and training; fp16 for GPUs without bf16
Notes
Shibari is VALIDATED, may allow obscure position support don't hesitate to try
中文：基于 Qwen Image 2.1，开箱即可生成裸体与成人场景，无需 LoRA。两个版本：Turbo（6 步，无 CFG）和 Base（完整模型，25–50 步，cfg 3）。8–12G 显存用 int8 convrot，6–8G 用 int4 mixed。仅限非商业用途。
License
Built with Qwen. Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026 Hangzhou Tongyi Laboratory Technology Co., Ltd. All Rights Reserved. Free, non-commercial use only. Full licence: https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE. LICENSE and NOTICE, with the list of changes, are in the zip.
Description
Turbo rebuilt in 6 steps.
In our blind test on the same 150 prompts and seeds as V2 Turbo, fewer images had a visible defect with the same prompt adherence and explicit rendering.
Five formats: int8 convrot, int4 mixed, fp8, bf16, fp16. The workflow uses core ComfyUI nodes only. Built with Qwen.
FAQ
Comments (31)
any chance of no turbo lora backed in version?
Edid: thank you for adding it. will test it now.
Enjoy! V3 Base is the one I recommend.
Did you use your example images above with RAW (not turbo) model? I ask because I am rendering images with your workflow, with turbo (5-steps), V20 turbo fp8 diffusion version, and I want to improve image quality. Some outputs have no proper body parts. There is no 'step' node. How to increase steps without a Ksampler node?
For the best quality, use the Base version: its workflow is in the files of that version and uses a normal KSampler, so you can set the steps there (25 by default). The Turbo is built for its few fixed steps.
Will you add Natural Breasts and also larger Areola and Nipples to the next version?
Everything looks the same as if the dataset was very narrow or similar.
Very low quality if used to edit image, it makes the image quality very low.
Strange. for me it's the opposite. while editing the quality is great, while generating the quality is very low.
@infernahermit846 Hehe, sounds like something between Attention Engine, Sampler, Scheduler choice
or downright "which program runs it" or GPU differences ...
Try LCM
For the best quality, go with V3 Base, the Turbos were more experimental.
I cannot get this model in Turbo to work correctly as the images turn out really bad. Images using Base are better. I am using Kubuntu 26.04 Linux, Rocm 7.1 and an R9700 and SwarmUI.
I am using the recommended settings. In my opinion, Krea 2 is far superior at this time.
It would be very helpful to include example prompts.
Agreed, Base does better: V3 Base is the recommended version, the Turbos were more experimental. There's now a Prompt examples section on the page.
Cool, but why not a lora/lokr?
A full checkpoint is simpler to use, no extra files to set up.
don't bother with the TURBO MODEL. it provides absolutely unusable images with missing body parts and sometimes its just a mess of jumbled up pixels.
The Turbo versions were more experimental, V3 Base is the recommended one.
I really appreciate what you're trying to do here, but edits with this model are unusable and broken. Your V1 couldn't do nsfw very well, but V3 can't generate usuable images at all.
(talking about Turbo model)
RTX4060 8GB+32GB
NoctQ_V3_base_fp16.safetensors OK??
I recommend int8, same quality, 2 times faster
no anus, sad
There are LORAs that improve vagina/anus, etc. Try to search for 'vagina' and put BASE MODEL 'qwen 2.1', for example.
it works well (base model without turbo) but changing faces to much in edits plus removing hairs from public areas
I feel that v3.0 has both worse image editing performance and poorer prompt adherence compared to v2.0. Personally, I think the v2.0 Base version is the best overall.
V3.0 base on edit changes the face too much. V2 is better keeping face consistency.
V2 base and redqw21 are still the best tho for nsfw qwen2.1. But most loras with redqw21 are not compatible or half compatible to be used along with the model. Yours are compatible with all qwen21 loras i used.
3.0 works pretty well, haven't run into it changing faces. Model MUST be used in base form, turbo is pretty unusable sadly, But works great paired with the turbo lora here on civitai uploaded by tsolful.
The int8 version gives quite low quality text to image. I only tested v3 experimental at 1024x1024.
In editing gave several photos and told to convert to cartoon/ illustration style, but failed.
So I think only the regular version is good?
The Base version dose work on the above situations. Thanks for the Base model.
I'm flabbergasted in how better this model is compared with that guy that asked 88k buzz for it's model...
You'll be a lot less flabbergasted if you saw how often it happened with every other model family, the people asking absurd amounts of money in paygating and early access are never the highest effort model trainers/mergers. Many such cases.