Guides

Qwen Image Edit 2511: What Changed Since 2509 and How to Run It

What Qwen-Image-Edit-2511 changed over 2509, its Apache 2.0 license, official Diffusers settings, and where creators can post the edits.

By Anna Krejcova — Editor at FansCreatorPublished 11 min read
Article summary
  • Qwen-Image-Edit-2511 came out on 23 December 2025 under Apache 2.0. Qwen lists five gains over 2509: less image drift, better character consistency, built-in LoRAs, industrial design and geometric reasoning.
  • Qwen-Image-Edit-2509, from 22 September 2025, added multi-image editing (best with 1 to 3 images), steadier faces, products and text, and native ControlNet inputs. The original Qwen-Image-Edit, from 18 August 2025, takes one image.
  • 2509 and 2511 use the same Diffusers pipeline and settings: QwenImageEditPlusPipeline, 40 steps, true_cfg_scale 4.0, guidance_scale 1.0. The 2511 weights total about 58 GB.
  • For creators, 2511's strength is editing one person's outfit, scene, pose or lighting while the face stays the same. Platform AI rules still decide where the result can be posted.

Qwen Image Edit 2511 is the December 2025 update of Alibaba Qwen’s open-weight image editing model, published as Qwen-Image-Edit-2511. It came out on 23 December 2025 under the Apache 2.0 license, and its main gains over Qwen-Image-Edit-2509 are steadier faces, less image drift and selected popular LoRAs built into the base model. Qwen-Image-Edit-2509, from 22 September 2025, was the release that added editing with several input images at once.

Model facts below come from the QwenLM GitHub repository, the Hugging Face model cards in the Qwen organization and the Qwen-Image technical report, all read on 6 October 2026.

What is Qwen-Image-Edit?

Qwen-Image-Edit is the editing version of Qwen-Image, Alibaba Qwen’s image model. You give it a picture and a written instruction, and it returns the edited picture. The weights are open, so other services host it too. FansCreator describes its own Qwen setup on its Qwen Image Edit NSFW page. The original model card says it is “built upon our 20B Qwen-Image model”.

The model reads your picture twice. It sends the image to Qwen2.5-VL, a vision-language model, “for visual semantic control”, and to a VAE encoder “for visual appearance control”. The technical report calls this a “dual-encoding mechanism” that balances “preserving semantic consistency and maintaining visual fidelity”.

In practice that gives two kinds of edit:

  • Appearance edits change one element and leave the rest of the picture untouched, such as adding, removing or recoloring an object.
  • Semantic edits allow most pixels to change while the subject stays recognizable, such as a new style, a new pose or a rotated view.

It also edits text inside images, in English and Chinese, while keeping the original font, size and style.

Qwen-Image-Edit release timeline: base model and edit releases

Qwen ships the editing model as dated updates rather than new names. Qwen calls 2509 the September 2025 “monthly iteration” of the original, and the 2511 weights were published on 23 December 2025.

Date Release What it brought
4 August 2025 Qwen-Image The 20B text-to-image base model
18 August 2025 Qwen-Image-Edit Single-image editing, text editing in English and Chinese
22 September 2025 Qwen-Image-Edit-2509 Multi-image editing, better person, product and text consistency, ControlNet inputs
23 December 2025 Qwen-Image-Edit-2511 Less image drift, better character and group consistency, built-in LoRAs

All dates come from the News section of the GitHub README, which listed no newer edit model after 2511 as of 6 October 2026. On 19 August 2025, one day after the first release, the team posted that it had “observed performance misalignments”. It asked users to update to the latest Diffusers code, with gains “especially in identity preservation and instruction following”. If you tested the original model in its first days, your results may not match the fixed version.

What changed in Qwen-Image-Edit-2509

Qwen-Image-Edit-2509 is the “monthly iteration” of the original. Its model card lists three changes:

  • Multi-image editing. The model was “further trained via image concatenation” so it can combine pictures, for example “person + person”, “person + product” or “person + scene”. The card says “optimal performance is currently achieved with 1 to 3 input images.”
  • Better single-image consistency. It keeps facial identity better through new portrait styles and poses, keeps product identity for poster work, and can change the font, color and material of text as well as the words.
  • ControlNet inputs. It accepts depth maps, edge maps and keypoint maps natively, so a pose skeleton can set a new pose.

When 2509 shipped, the README recommended it over the original, saying 2509 “has better consistency than Qwen-Image-Edit” and should be used “for both single image input and multiple image inputs”. The original model takes one image only.

What changed in Qwen-Image-Edit-2511

The Qwen-Image-Edit-2511 model card calls it “an enhanced version over Qwen-Image-Edit-2509” and lists five improvements: “mitigate image drift, improved character consistency, integrated LoRA capabilities, enhanced industrial design generation, and strengthened geometric reasoning ability.”

  • Less image drift. Edited pictures stay closer to the input. The card names this gain but does not define it further.
  • Character consistency. The model “can perform imaginative edits based on an input portrait while preserving the identity and visual characteristics of the subject.” It also “further enhances consistency in multi-person group photos”, including fusing “two separate person images into a coherent group shot”.
  • Built-in LoRAs. It “integrates selected popular LoRAs directly into the base model”, so effects that needed an add-on before work without one. The examples are realistic lighting control and generating new viewpoints of a subject.
  • Industrial design. Batch product design and material replacement on industrial parts.
  • Geometric reasoning. It can draw auxiliary construction lines for design or annotation work.

The card does not put a number on any of these gains, so treat them as Qwen’s description rather than a measured benchmark.

Qwen Image Edit 2511 vs 2509 vs the original

Based on what each official model card states, read 6 October 2026:

Qwen-Image-Edit Qwen-Image-Edit-2509 Qwen-Image-Edit-2511
Released 18 August 2025 22 September 2025 23 December 2025
Input images One Several, best with 1 to 3 Several (the card example uses two)
Diffusers pipeline QwenImageEditPipeline QwenImageEditPlusPipeline QwenImageEditPlusPipeline
Steps in the example 50 40 40
true_cfg_scale 4.0 4.0 4.0
guidance_scale Not set 1.0 1.0
Negative prompt A single space A single space A single space
License Apache 2.0 Apache 2.0 Apache 2.0
Main strength Text editing, semantic edits Combining several pictures Holding one face steady, group shots

2509 and 2511 share the same pipeline class and the same settings, so moving from one to the other is mostly a change of model name.

Qwen-Image-Edit-2511 license

All three editing models list Apache 2.0 on their Hugging Face cards, and the GitHub README states “Qwen-Image is licensed under Apache 2.0.” Apache 2.0 is a permissive license. It allows commercial use, changes and redistribution, as long as you keep the license and its notices with what you share.

The license covers the code and the weights. It does not cover the people in your pictures or the rules of the site where you post.

How to run Qwen-Image-Edit-2511

1. Try it online

Two official options need no install. Qwen Chat has an Image Editing feature, which the model card points to for “the latest model”. Qwen also hosts a Hugging Face demo for 2511.

2. Run it in Python with Diffusers

This is the route the model card documents.

  1. Install Diffusers from source: pip install git+https://github.com/huggingface/diffusers. The README also asks for transformers 4.51.3 or newer, for Qwen2.5-VL support.
  2. Load the pipeline in bfloat16 and move it to a CUDA graphics card.
  3. Pass a list of input images, your instruction and the settings from the table above.

The loading step from the official example:

pipeline = QwenImageEditPlusPipeline.from_pretrained(
    "Qwen/Qwen-Image-Edit-2511", torch_dtype=torch.bfloat16
)
pipeline.to("cuda")

The model card gives no minimum video memory figure. What it does show is size. The weight files in the 2511 repository add up to about 58 GB as of 6 October 2026. About 41 GB is the image transformer and about 17 GB is the Qwen2.5-VL text encoder. Plan disk space and memory for that. For a lighter Alibaba model, the Z-Image Turbo guide covers one whose README says it fits within 16 GB of video memory.

3. Use another inference engine

The README lists day-0 support for 2511 from three open-source serving projects: vLLM-Omni, SGLang-Diffusion and LightX2V. LightX2V reports “an order-of-magnitude 42.55x overall speedup” through distillation. That figure is the project’s own claim, quoted in Qwen’s README, not an official Qwen benchmark.

How to write a Qwen-Image-Edit instruction

Describe the change, not the whole picture. With several inputs, say which picture is which. The official examples do it two ways: by position (“The magician bear is on the left, the alchemist bear is on the right”) or by number, referring to each input as Figure 1, Figure 2 and so on.

What Qwen-Image-Edit-2511 is good at for creators

Consistent edits of one person. Change the outfit, the room or the pose of a portrait while the face stays the same. That is the “character consistency” gain, and it is why 2511 suits a feed built around one recognizable face. Wardrobe and scene changes on a persona are covered in virtual model wardrobe and scene swapping.

Multi-image editing. Put a person into a scene from another photo, or next to a product, in one edit. 2511 also holds faces better when two separate portraits are merged into one group shot.

Lighting and new angles. The built-in LoRAs mean relighting a shot or showing the subject from a new viewpoint works without hunting for add-ons.

Text on images. Promo graphics, signs and posters keep their typography when you change the words.

There are limits too. The model only knows the images you give it in that edit, so it does not remember a person between sessions. For a whole library of one face, compare the methods in LoRA training vs multi-angle prompting.

Before you post edited pictures

Only edit pictures of your own face, your own AI persona, or a person whose consent you can prove. Use your own original photos as references, ideally ones you have never posted to OnlyFans. The OnlyFans Terms of Use define Content as “any material Uploaded to OnlyFans by any User” and say you may not use Content “to train or enhance any software, machine learning or artificial intelligence system”. The FansCreator Terms of Use prohibit anything else, and section 5.2 asks for a Consent Record even when the person in the picture is you.

Then check the rules of the site you post on:

  • OnlyFans allows AI content only if “it clearly features the verified Creator” and carries a label such as #ai or #AIGenerated, per its Terms of Use and Help Centre. Its removal guide lists content that is “wholly AI-generated” or “features heavily altered images”. A label alone does not make content allowed. The full clauses are on our OnlyFans AI content policy page.
  • Fanvue allows fully AI-generated creators with a clear AI disclosure in the bio, a caption or a watermark. Content that is “deepfaked/face-swapped” is allowed too, but “you MUST verify the real model behind the AI” (Fanvue Help Centre).
  • MYM allows AI content labeled “Virtual Image”, #AI or #AIgenerated, and adds an “AI profile” badge. Deepfakes are allowed only if “you have written permission from the person being deepfaked” (MYM support).
  • Fansly bans photorealistic AI content, including face swaps, even when labeled (Fansly Help Centre).

So light edits that clearly show you can go to OnlyFans with a label; heavily altered images can still be removed even when labeled. An original persona belongs on Fanvue or MYM. If you are starting that persona, how to create an AI influencer walks through it, and AI pictures of yourself covers the OnlyFans side.

Which Qwen-Image-Edit version to use

Use Qwen-Image-Edit-2511 for new work. It has the same settings as 2509, keeps faces steadier and handles group shots. Stay on 2509 only if a workflow or LoRA you depend on was built for it. Keep the original Qwen-Image-Edit for reproducing old results, since it takes one image and Qwen’s README recommends 2509 over it.

If you would rather not host the model yourself, FansCreator’s uncensored AI image editor edits a picture from a plain-language instruction in the browser. It keeps one creator’s face the same across images. You download the result and post it yourself.

Frequently asked questions

What is the difference between Qwen-Image-Edit-2511 and 2509?

Qwen-Image-Edit-2509 (22 September 2025) added multi-image editing, better person, product and text consistency, and ControlNet inputs. Qwen-Image-Edit-2511 (23 December 2025) builds on it with less image drift, better character consistency including group photos, selected community LoRAs built into the base model, industrial design work and stronger geometric reasoning. Both use the same Diffusers pipeline and settings.

Can I use Qwen-Image-Edit-2511 commercially?

The Hugging Face model card and the GitHub repository list the Apache 2.0 license, which allows commercial use, changes and redistribution if you keep the license and notices. The license does not cover the people in your pictures or the rules of the site where you post.

How many images can Qwen-Image-Edit-2511 take?

More than one. The 2509 model card says optimal performance is achieved with 1 to 3 input images, and the 2511 card's example passes two images. The original Qwen-Image-Edit accepts a single image only.

What settings does Qwen-Image-Edit-2511 use?

The official example loads QwenImageEditPlusPipeline in bfloat16 on a CUDA GPU and uses 40 inference steps, true_cfg_scale 4.0, guidance_scale 1.0 and a single space as the negative prompt.

How big is Qwen-Image-Edit-2511?

The weight files in the official Hugging Face repository add up to about 58 GB as of 6 October 2026: about 41 GB for the image transformer and about 17 GB for the Qwen2.5-VL text encoder. The model card does not give a minimum video memory figure.

Sources

Read as Markdown

Bring your next idea to life.

Create images, videos and content for your audience with FansCreator.

Start for Free