Qwen Image Edit: Versions, Licence and ComfyUI Setup

Qwen-Image-Edit is an open image editing model from the Qwen team at Alibaba Cloud: you give it a picture and a written instruction, and it returns the edited picture. The original model and its two updates, 2509 and 2511, are licensed under Apache 2.0 and run natively in ComfyUI from three files.

Checked on 13 October 2026 against the model cards for Qwen-Image-Edit, 2509, 2511, Qwen-Image-2.1 and 2.1-Turbo, the Qwen-Image repository and its licence file, the Qwen Research License, ComfyUI's Qwen-Image-Edit, 2511 and Qwen-Image-2.1 tutorials, its Qwen examples, and Comfy Org's repackaged files for the editing models and the text encoder and VAE. We have not timed or scored the models ourselves.

Quick answer

What is Qwen-Image-Edit, and what is Qwen?

Qwen is, in the words of its Hugging Face organisation page, "the large language model family built by Alibaba Cloud". The same team publishes image models under the Qwen-Image name. Qwen-Image is the text-to-image model, which the repository describes as a 20B MMDiT image foundation model, and Qwen-Image-Edit is its editing version.

The model card explains how it edits. The input image goes to two places at once: to Qwen2.5-VL, a vision language model, for what the card calls visual semantic control, and to the VAE encoder for visual appearance control. That gives two kinds of edit:

• Appearance editing adds, removes or changes one element while the rest of the image is meant to stay unchanged. The card's examples include changing a background, changing clothing and adding a signboard.
• Semantic editing lets the whole image change while the subject stays the same, as in style transfer or rotating an object to a new viewpoint.
• Text editing adds, deletes or rewrites text inside an image in Chinese or English, which the card says preserves the original font, size and style.

The card also claims state-of-the-art results on public editing benchmarks. That is the publisher's claim, not our ranking. New to node workflows? Start with what ComfyUI is, then see the other open models in best ComfyUI models.

Which Qwen image editing models exist today?

Release dates are from the news lists in the Qwen-Image and Qwen-Image-2.1 repositories; features and licences are from each model card. One more name appears there: Qwen-Image-2.0, announced on 10 February 2026 and offered through Qwen Chat. The Qwen organisation on Hugging Face listed no Qwen-Image-2.0 weights when we checked.

Is Qwen-Image-Edit free to use commercially?

For the three Qwen-Image-Edit models, yes. Each model card carries the Apache 2.0 tag and the line "Qwen-Image is licensed under Apache 2.0", and the repository's licence file is the Apache License, Version 2.0. Apache 2.0 permits commercial use and asks you to keep the licence and notices when you redistribute the weights or code. The weights are free; the graphics card or the rented GPU time is not.

Check the licence tag on the exact repository you download: two models from one team now sit under different terms.

How to run Qwen-Image-Edit in ComfyUI

ComfyUI supports the Qwen-Image family natively, and its official tutorial ships a native workflow in the Template Library, so the basic edit needs no custom nodes.

File names and folders are from ComfyUI's tutorial; sizes are from Comfy Org's Hugging Face repositories and, for the LoRA, from lightx2v's repository. The three required files come to about 30 GB. Two details from the tutorial save time:

• The resize node is there on purpose. It avoids the quality loss the tutorial describes with oversized inputs such as 2048x2048. If you know your image size, bypass it with Ctrl+B.
• The 4-step LoRA is off by default. Select the LoraLoaderModelOnly node and press Ctrl+B to enable it when you want faster drafts.

The 2509 and 2511 files

The newer versions reuse the same text encoder and VAE, so you only swap the diffusion model. ComfyUI's examples page uses qwen_image_edit_2509_fp8_e4m3fn.safetensors (20.4 GB) and notes that 2509 supports up to 3 image inputs. The 2511 tutorial lists qwen_image_edit_2511_bf16.safetensors, a 40.9 GB file, and the same repository holds 20.5 GB versions. Qwen-Image-2.1 uses different files throughout, listed in its own tutorial.

If you add community node packs later, install them through ComfyUI Manager and keep the official template as the version you know works.

What hardware does Qwen-Image-Edit need?

The Qwen model cards give no minimum graphics card, and neither does ComfyUI's tutorial. What you can plan around is published file size: a 20.4 GB diffusion model and a 9.38 GB text encoder in fp8, where Comfy Org also lists a 6.11 GB version of the text encoder. ComfyUI's startup flags reference says weight offloading is on by default on Nvidia and AMD graphics cards, with dynamic VRAM enabled where it is supported. Offloading lets a model larger than your graphics memory run, more slowly.

Outside ComfyUI, the Qwen repository says the DiffSynth-Studio library provides "low-GPU-memory layer-by-layer offload (inference within 4GB VRAM)", a statement about one library with no speed attached. We give no generation times because no primary source publishes them for this model. ComfyUI's own documented requirements are on our ComfyUI statistics page.

How to write a Qwen-Image-Edit prompt

Write an instruction, not a description of a finished picture. The card's own example is one sentence: "Change the rabbit's color to purple, with a flash light background." Name the thing that changes, then say what must stay.

• Outfit colour: "Change the dress to dark green. Keep her face, hair, pose and the background unchanged."
• The same edit in Chinese: "把裙子颜色改为深绿色，保持脸部和姿势不变"
• Background: "Replace the background with a beach at sunset. Keep the subject and her lighting the same."
• Text on a sign: "Change the sign text to 'NEW STORE' in the same font and colour."
• Add an object: "Add a potted plant on the table to the left, matching the light in the room."

Hard edits go in steps. The card corrects a calligraphy image one character at a time, feeding each result back in as the next input, and calls this chained editing. To draft the first still of a new persona, our free AI influencer prompt generator writes a structured starting prompt.

Whose face are you editing?

An image editor changes a person in a photo as easily as a sign. A persona is a fictional adult you created, or you. Editing or reproducing a real person's face without their written permission is not allowed, and that covers swapping an outfit or a background on a photo of someone else. Apache 2.0 licenses the weights and code and has no clause about likeness, so the limits come from the law where you live and the rules of the platform you post on.

Qwen-Image-Edit vs FLUX.1 Kontext

The other open instruction editor most people compare is FLUX.1 Kontext [dev] from Black Forest Labs, covered in our FLUX Kontext guide. These rows are each publisher's own statements, plus ComfyUI's documentation and file listings. Neither column is a test result.

Kontext facts are from its model card and ComfyUI's Kontext Dev tutorial. The licence row decides most cases: for business images, Qwen-Image-Edit runs locally under Apache 2.0, while Black Forest Labs asks for a commercial licence to run Kontext [dev]. If you prompt in Chinese, Qwen documents that directly. For your persona, test both on the same still and judge the faces yourself.

Qwen LoRA training: what the sources say

A LoRA is a small add-on file that teaches a base model one subject or style. The Qwen repository says three things about them. The Qwen-Image family has supported LoRA models since 9 August 2025. DiffSynth-Studio supports "LoRA / full training" for Qwen-Image, and ModelScope hosts a LoRA training page, while LoRA and fine-tuning support in Diffusers is described as in development. And Qwen-Image-Edit-2511 integrates selected popular community LoRAs into the base model, so their effects work without loading a file.

The repository gives no recommended image or step count for a character LoRA, so we quote none. Dataset basics are in our LoRA training guide, and the free LoRA training steps calculator turns images, repeats and epochs into a step total. Train only on a fictional persona or on yourself. The AI Influencers program teaches the full pipeline, and the free guides are grouped in the AI influencers library.

Qwen-Image-Edit: FAQ