Qwen-Image 2.1, uncensored on your own card
Qwen-Image 2.1 is the image model Alibaba Qwen published on 20 September 2026, and the first in its series where one set of weights both generates a picture from a prompt and edits a reference image. Locally Uncensored 3.0.1 carries it on Windows and Linux as a Model Manager bundle called Qwen-Image 2.1 (Generate and Edit): three files, 16.1 GB, straight into the right ComfyUI folders. After that it is an entry in the Image lane picker and the Edit lane picker, like any other local image model.
It needs ComfyUI 0.37.0 or newer. If the node is missing, the app says so in one sentence and points at the update button.
Three ways in
How to Run Qwen-Image 2.1 Locally
The step by step version: the three files and their sizes, the memory guide, the ComfyUI version check, the Image lane, the maskless Edit lane, and what this release does not do.
Qwen-Image 2.1 Explained
7B where the first Qwen-Image had 20B, a VAE with an alpha channel, ten reference images in a pass, no separate edit model, and a license that is no longer Apache 2.0.
Handbook: Qwen-Image 2.1 in the Create chapter
The short reference version, in the chapter that covers every lane: where the bundle is, which picker it appears in, what the Edit lane takes, and the message you get on an older ComfyUI.
What uncensored means for Qwen-Image 2.1
Two different things get called uncensored here, and it helps to keep them apart.
Where it runs. On your own machine nothing sits between your prompt and the model. The content policy setting belongs to LU Cloud. It applies to images and video rendered in the cloud, and nothing on your own machine is touched by it.
The weights. The original model draws what it was trained to draw. Community builds labelled uncensored change the weights or the text encoder. We list them below as their authors describe them. We have not tested them.
Uncensored builds and GGUF files
File sizes were read off the repositories on 10 October 2026. Both GGUF repositories name the Qwen Research License.
| Repository | What it is | Files |
|---|---|---|
| abenzerps/Qwen-Image-2.1-Uncensored-GGUF | Uncensored build, published 20 September 2026 | Q4_K_M 4.6 GB, Q5_K_M 5.22 GB, Q6_K 5.88 GB, BF16 14.23 GB |
| unsloth/Qwen-Image-2.1-GGUF | The original weights as GGUF | Q2_K 2.47 GB, Q4_K_M 4.2 GB, Q5_K_M 5.39 GB, F16 14.23 GB |
| pottokao/Qwen-Image-2.1-Text-Encoder-Heretic-GGUF | The text encoder only, processed with Heretic | Q4_K_M 5.03 GB, Q6_K 6.88 GB, Q8_0 8.71 GB |
The bundle in Locally Uncensored installs the original int8 files, three files and 16.1 GB. The files in the table go into ComfyUI by hand, and ComfyUI reads GGUF through a loader node that has to be installed first. For what abliteration and Heretic do to a model, see the abliterated models guide.
The uncensored builds inside the app
Since 3.0.5 two of these come as downloads under Models, Image, Get new, so nothing has to go in by hand:
| Download | What it changes | On top of an existing install |
|---|---|---|
| Qwen-Image 2.1 (No Filter) | The text encoder. Same image model and VAE, with a community text encoder that has its refusal direction removed (Heretic) | About 9 GB, only the encoder |
| Noct Q (Qwen-Image 2.1, Unfiltered) | The image model. A community finetune by Noctaluna whose notice says the transformer weights were modified. It uses the official text encoder and VAE | About 7 GB, only the image model |
With both text encoders installed, the Expert settings in Create have a "Text encoder" row to pick which one reads your prompt, for new pictures, for edits and for pictures made in chat. The official one stays the default. The app treats Noct Q as Qwen-Image 2.1, so edits with reference images, the transparent background, the encoder choice and the prompt enhancer work with it. We have not run Noct Q on our own hardware yet. Both carry the Qwen Research License.
The prompt enhancer comes the same way, as two downloads, Official and No Filter. Both run on a 12 GB card and need ComfyUI 0.37.2 or newer.
LoRAs for Qwen-Image 2.1
When 3.0.1 shipped there was no LoRA for 2.1, and the older Lightning and Turbo LoRAs do not fit the new architecture. That has changed. On 10 October 2026 Hugging Face listed these, among others:
- Viggle/Qwen-Image-2.1-viggle-turbo, speed LoRAs for four to six steps, 0.34 GB to 1.36 GB
- alibaba-pai/Qwen-Image-2.1-Fun-Acc-LoRAs, acceleration LoRAs, published 24 September 2026
- f23gg/NSFW-LORA-Qwen-Image-2.1, a content LoRA of 0.16 GB, tagged not for all audiences by Hugging Face
The LoRA stack is open for every local image model. Drop the .safetensors file into ComfyUI's models/loras folder and hit Rescan. We have not tested these LoRAs with the bundle.
The license, before you download 16 GB
What the app does with it today
It generates from a prompt with a negative prompt, at the defaults from the official template: 25 steps, CFG 1, euler, simple, 1024x1024, with the usual Advanced settings for Steps, CFG, Width, Height, Seed and Batch. It edits from a reference image plus a prompt, with no mask. Small tiles under the source image take up to three more images, which you name in the prompt as image 2 and image 3, for example "put the jacket from image 2 on the person". A "Transparent background" switch saves a new picture as a PNG with an alpha channel, and "Improve my prompt" can use the Qwen-Image 2.1 prompt enhancer.
Still not there: masks with this model and the KV cache node. The older Lightning and Turbo LoRAs do not fit the new architecture.
On a Mac
The Mac build for Apple Silicon (beta, since 3.0.6) runs Qwen-Image 2.1 without ComfyUI, as quantized packs. The 4-bit pack needs 16 GB, the 8-bit and Turbo packs ask for 24 GB. Those are the minimums in the catalog; the build has not run on a Mac with 16 or 24 GB yet. On an M5 Pro with 48 GB, the only Mac the build has run on, Qwen-Image 2.1 Turbo made an image in 38 seconds. Edit on the Mac works by prompt, without a mask. A LoRA that carries its own alpha is weighted alpha divided by rank, as in ComfyUI. The Mac page has the state of the build.
FAQ
Is Qwen-Image 2.1 uncensored?
The original weights draw what they were trained to draw. On your own machine nothing sits between your prompt and the model, and community builds labelled uncensored have been on Hugging Face since 20 September 2026. We list them as their authors describe them and have not tested them.
Is there a GGUF of Qwen-Image 2.1?
Yes. unsloth/Qwen-Image-2.1-GGUF carries the original weights from 2.47 GB at Q2_K to 14.23 GB at F16, with Q4_K_M at 4.2 GB. The bundle in Locally Uncensored installs the original int8 files instead, three files and 16.1 GB.
Are there LoRAs for Qwen-Image 2.1?
Yes, since late September 2026. Speed LoRAs with four to six steps and content LoRAs are on Hugging Face. The LoRA stack in Locally Uncensored lists every .safetensors file in ComfyUI's models/loras folder.
Is there an uncensored Qwen-Image 2.1 in Locally Uncensored?
Two downloads in the Model Manager. Qwen-Image 2.1 (No Filter) swaps the text encoder for one with its refusal direction removed, about 9 GB on top of an existing install. Noct Q (Qwen-Image 2.1, Unfiltered) is a community finetune that changes the image model itself, about 7 GB on top. Both carry the Qwen Research License.
Can I use Qwen-Image 2.1 commercially?
Not under the published terms. Qwen-Image 2.1 is under the Qwen Research License, which allows research and evaluation. Commercial use needs a separate license from Alibaba Qwen.
Locally Uncensored is free and AGPL-3.0, for Windows, Linux and Macs with Apple Silicon (beta). New here? The handbook takes you from installer to first image. The rest of the family is under all Qwen guides.