How to Run Qwen-Image 2.1 Locally: One Model That Draws and Edits
Until now, drawing a picture and editing one meant two sets of weights on your disk. Qwen-Image 2.1 does both jobs with the same files, and since Locally Uncensored 3.0.1 it is a one click download in the Model Manager on Windows and Linux. This guide walks through the download, the ComfyUI version it insists on, the two lanes it appears in, and the license you should read before you plan any work around it.
What Qwen-Image 2.1 is, in three sentences
Qwen-Image 2.1 is the image model Alibaba Qwen published on 20 September 2026, and it is the first in the series that generates from a prompt and edits a reference image with one set of weights, with no separate edit model beside it. The image part is a 7B single stream DiT with 32 layers, the text encoder is Qwen3-VL-8B, and the VAE is a new 64 channel design that can carry an alpha channel. The predecessor, the August 2025 Qwen-Image, used a 20B image model and Qwen2.5-VL, so 2.1 is a smaller model rather than a larger one.
What 2.1 adds over everything before it: a native alpha channel, up to ten reference images in a single pass, local edits driven by a drawn marking instead of only a mask, a prefix KV cache that speeds up editing, and better typography. Native 2K rendering carried over from 2.0 in February 2026. Not all of that reaches the app yet, and the honest list of what does not is further down this page.
What it wants from your machine
We have not measured this model ourselves, so the numbers here are quoted, not claimed. As a guide, ai.rs puts the int8 pipeline at roughly 16 GiB in total: about 6.8 GiB for the image model, 8.7 GiB for the text encoder and 0.6 GiB for the VAE. A card with 24 GB holds all three parts resident at once. On a 16 GB card the usual move is to keep the text encoder in system RAM and let the card carry the rest. The full bf16 weights are another class of machine entirely at 33 to 40 GB.
This guide covers the Windows and Linux builds of the desktop app. A Mac build exists since version 3.0.6, for Apple Silicon, and it is not part of this guide.
The three files
The bundle is the int8 repack from Comfy-Org, three files that land in three different ComfyUI folders:
| File | Folder | Size |
|---|---|---|
qwen_image_2.1_int8_convrot.safetensors | models/diffusion_models/ | 7.26 GB as Hugging Face lists it, 6.76 GiB measured |
qwen3vl_8b_int8_convrot.safetensors | models/text_encoders/ | 9.35 GB listed, 8.71 GiB measured |
qwen_image_2.1_vae_bf16.safetensors | models/vae/ | 676 MB listed, 0.63 GiB measured |
The Model Manager reports the bundle as 16.1 GB. Full bf16 versions of the first two files exist at 14.2 GB and 17.5 GB, and there is a w4a8 variant of the encoder at 6.31 GB; Locally Uncensored uses neither, it installs the int8 set.
Step 1: get 3.0.1
Qwen-Image 2.1 arrived with Locally Uncensored 3.0.1, which has been the Latest release since 22 September 2026. Download it from GitHub Releases, one installer for Windows and three packages for Linux. If you are already on a 3.0.x build, the signed update channel picks it up at the next launch. The install chapter of the handbook lists the file names if you want to check what you downloaded.
Step 2: install the bundle in the Model Manager
Open the Models tab and find the bundle named Qwen-Image 2.1 (Generate and Edit). Its description names the license before you start: Qwen Research License, non-commercial use. One click downloads all three files into the right ComfyUI folders, and the app counts the bundle as installed once the running ComfyUI actually lists the files. That last detail matters: if ComfyUI is stopped, the rails cannot report what is on disk, so start it first.
Step 3: make sure ComfyUI is 0.37.0 or newer
Support for this model landed in ComfyUI 0.37.0, merged on 19 September 2026, and it brought a new encode node, TextEncodeQwenImage21. The old Qwen-Image edit nodes do not fit. If your ComfyUI is older, Locally Uncensored says so in one sentence rather than failing halfway through a graph:
Do what it says: Settings, AI Backends, ComfyUI, then "Update ComfyUI". The output panel under the buttons shows what happened. Once the new node is registered, the model becomes selectable.
Step 4: generate in the Image lane
Open Create. The Image lane is the default one. Pick Qwen-Image 2.1 in the model picker; there is no new button and no new tab for it, it sits in the same picker as every other local image model. Write a prompt in the composer, add a negative prompt if you want one, and press Create.
The defaults follow the official template: 25 steps, CFG 1, the euler sampler, the simple scheduler, 1024x1024. CFG below 1 is not usable with this model, so if you go exploring under "Advanced settings", leave that dial alone or raise it rather than lower it. The advanced panel is the same one the other lanes use: Steps, CFG, Width, Height, Seed and Batch size.
Step 5: edit in the Edit lane
Switch to the Edit lane, pick Qwen-Image 2.1 in the picker there, drop in a reference image, and describe the change you want. Since 3.0.7 the strip under it takes up to nine more images, so ten in all; from four extra images on, the hint asks you to say what each one brings, for example "the scarf from image 3". There is no mask to paint. The mask path is deliberately closed for this model, so the output follows the reference image and the prompt, and the resolution follows the canvas you picked.
If you are used to painting a red area in this lane with an SDXL checkpoint, that habit does not transfer here. It is a different way of working, not a broken one: you say what should change in words instead of drawing where.
The license, stated plainly
This is the one point worth being blunt about, because the rest of the series trained people to expect Apache. The bundle description in the Model Manager carries the same sentence, so nobody downloads 16 GB and finds out afterwards.
What is not in this release
Locally Uncensored 3.0.1 does not expose everything the model can do. Stated as it stands today, with no promises attached:
- Multiple reference images. The model takes up to ten in one pass. The Edit lane took one in 3.0.1. Update: since 3.0.7 it takes up to ten too. With more images than fit at full size, every image is scaled smaller so that together they need no more memory than the full size ones would: four on larger cards, three on a 16 GB card and two on a card with less than 16 GB. With four, ten images run at 640 pixels each at a 1024 canvas, and on ComfyUI the result comes out at that smaller size as well. Ten images have not run with a real model on Windows and Linux yet. On a Mac, 3.0.7 also adds masks and a prompt enhancer for this model.
- RGBA output. The new VAE carries an alpha channel. The app does not write one out. Not in this release.
- Masks with this model. Closed on purpose, as described above.
- A prompt enhancer and the KV cache node. Neither is wired up. Not in this release.
- A speed LoRA. There is no Lightning or Turbo LoRA for 2.1 as of 21 September 2026, and the old ones do not fit the new architecture. The LoRA stack stays open for every local image model, so there is simply nothing to put in it.
And one thing we are not going to put a number on: how long a render takes. Nobody here has timed this model on a card, so any second count you read here would be someone else's. ai.rs reports 5.93 seconds on an RTX 5090, which is their figure on their hardware, not ours.
Download
Locally Uncensored 3.0.1 is free and AGPL-3.0, on GitHub Releases. The Create chapter of the handbook covers the lanes this model joins, and the Qwen-Image 2.1 page collects everything we have written about it.
Related
- Qwen-Image 2.1 explained, the architecture, the timeline and the license change
- v3.0.1 release notes, the release this model arrived in
- The easiest local AI image generator, the Create tab from the beginning
- Image to image with local AI, the lane that now has a maskless option
Locally Uncensored 3.0.1. AGPL-3.0 licensed. Built by PurpleDoubleD. Bug reports and feature requests on GitHub Issues or in the Discord.
In the handbook
- All Qwen guides, every generation on one page
- How to run FLUX locally, the step by step for the image model most people start with
- Create: images, video, audio and characters, image generation and editing, video, music and speech
- First start: the setup wizard, every step of the wizard, from the engine to the first model