Qwen-Image-2.1

Qwen-Image-2.1 generates images with a real alpha channel, which almost nothing else in the open image-generation world does. It edits from up to ten reference images, takes circles drawn on a picture as instructions, and its ComfyUI repack has been pulled over 3.6 million times. It is also research-only. Before you build anything on it, read the license, because "open weights" is doing a lot of work in this release.

Transparency is the feature

Every image model can make a picture of an object on a white background. Getting that object onto a transparent background has meant a separate background-removal step, a matting model, and a fight with hair and glass edges. Qwen-Image-2.1 generates RGBA natively, alpha channel included, which removes that step entirely.

You ask for it in the prompt rather than through a flag. Qwen's recommended phrasing is explicit: "This is an RGBA image with transparency... The image has alpha channel and the background is transparent." That is a slightly odd interface, since it means the capability depends on hitting the right words, but it works without extra tooling.

If you produce UI assets, product cut-outs, sprites, overlays or anything destined for a compositing step, this is the practical reason to care. It removes a stage from the pipeline rather than improving one.

What else is new

The visual generation component is 7B parameters across 32 single-stream DiT layers, with mixed-granularity attention and prefix KV cache reuse for efficiency. Qwen positions creation and editing as one unified capability rather than separate models.

Editing takes up to ten reference images at once, which is well beyond the one or two most editing models accept, and it preserves identity for both people and products across edits. Local edits can be directed by drawing on the image: circles, painted annotations or masks all work as spatial instructions, so you can point instead of describing position in words.

Output resolutions are generous and cover the aspect ratios you actually ship: 2048x2048 at 1:1, 2400x1792 at 4:3, 2752x1536 at 16:9, 1536x2752 at 9:16, plus 3:2 and 2:3 variants. Qwen also claims improvements to typography, portrait lighting and fine detail. No benchmark numbers ship with the model card, so those quality claims are assertions rather than measurements.

The license is research-only

This is the part that decides whether the rest matters to you. Qwen-Image-2.1 is under the Qwen Research License Agreement, and the restriction is not subtle: "You shall not use the Materials for any commercial purpose without obtaining a separate commercial license from us." The agreement defines Non-Commercial as "for research or evaluation purposes only."

So a company cannot ship this in a product. A freelancer cannot use it for client work. Unlike the weights-versus-outputs split some licenses draw, this one restricts the use of the materials themselves, which is the broader reading.

Derivative works and fine-tunes are permitted within those bounds. If you distribute a model built on it you must display "Built with Qwen" or "Improved using Qwen", modified files must carry prominent change notices, and the copyright line for Hangzhou Tongyi Laboratory has to be retained.

It is worth lining this month's releases up, because the spread is wide and it is the thing that most often gets flattened into the phrase "open source":

Qwen-Image-2.1 sits at the restrictive end. That is Qwen's right, and the weights are genuinely downloadable and modifiable for research, but it is a different proposition from the Apache 2.0 terms most of the Qwen language models carry. Do not assume the text-model license applies here.

Running it

Diffusers support is in place through a dedicated pipeline, and CPU offloading is available if you are short on VRAM. Qwen does not publish a VRAM figure, so budget by experiment rather than by the spec sheet.

from diffusers import QwenImage21Pipeline

pipe = QwenImage21Pipeline.from_pretrained("Qwen/Qwen-Image-2.1")
pipe.enable_model_cpu_offload()

image = pipe(
    "This is an RGBA image with transparency. A ceramic coffee cup, "
    "studio lighting. The image has alpha channel and the background is transparent."
).images[0]
image.save("cup.png")

For most people the faster route is ComfyUI through the Comfy-Org repack, which is where the 3.6 million downloads are concentrated and where the editing workflows with reference images are easiest to wire up. Our ComfyUI and Flux comparison covers that setup, and the self-hosted image generator guide puts the VRAM tradeoffs of the alternatives side by side.

Who should use it

Use it if you are researching, evaluating, or building something you will never sell, and you want native transparency and multi-reference editing that nothing else open offers right now. For that audience it is the most capable open image model available this month.

If money will ever change hands, go back to Flux 2 or another permissively licensed option and accept the extra background-removal step, or contact Qwen about a commercial license. The transparency feature is good. It is not good enough to justify shipping under terms that forbid shipping.

Sources and further reading

• Qwen-Image-2.1 model card
• Qwen Research License Agreement, full text
• The Comfy-Org repack for ComfyUI

Ten minutes: open the license before you open the model card. If the research-only clause rules you out, you have saved yourself an afternoon of building a pipeline around something you cannot deploy, and that is the most valuable thing this release can tell you.

Tested on: not independently tested. Architecture, resolutions, capabilities and the recommended RGBA prompt format come from Qwen's model card; the license terms are quoted from the LICENSE file in the repository. Qwen publishes no benchmark numbers for this model, so the quality claims are the vendor's own and unmeasured. Download counts were read from the Hugging Face API on the date below.
Date checked: 2026-09-27