Qwen-Image-2.1 Ships RGBA Transparency and 2K Output, With a Licence That Bars Commercial Use

Alibaba's Qwen team published Qwen-Image-2.1 on September 20, 2026, and the number the model card leads with is the size: 7B parameters in a visual generator built from 32 single-stream diffusion transformer layers. The feature most likely to change a video maker's afternoon is smaller than that headline number. It is a 64-channel RGBA autoencoder, which puts a real alpha channel inside the generation instead of after it.

The transparent output is the point

Most image models return a rectangle. To place a logo, a product cutout or a sprite over a generated background you either generate against the background you want, or run a matting model afterwards and clean the edges by hand. Qwen-Image-2.1 generates the alpha channel in the same pass, and it does the reverse too: it edits a transparent layer, and it lifts a subject out of a photograph, so a character pulled from one image can be placed in another without leaving the model.

ComfyUI's write-up of its day-zero support says the same thing in pipeline terms: sprites, logos, icons and product cutouts come out of the sampler ready to composite, with no background-removal node, no matting model and no edge cleanup. For anyone assembling overlays, lower thirds, thumbnails or a character asset library, that is a step removed from every job.

What the model card lists

The specifications that matter for a working pipeline, as published:

• Native 2048x2048 output, generated at that resolution rather than upscaled into it. The card lists seven aspect ratios, among them 2752x1536 at 16:9 and 1536x2752 at 9:16.
• Up to 10 reference images in one pass, attached as image_1 onward in ComfyUI's text-encode node or fed to the Diffusers pipeline.
• Local edits specified by circle, painted annotation or a separate mask, with identity preservation for people and products.
• 40 inference steps, a Qwen3-VL 8B text encoder, and two 9B prompt-rewriting checkpoints shipping alongside the main weights.
• Qwen's own figures describe a 2K edit with ten reference inputs finishing in about 1.59 seconds on its reference hardware, as reported by Pondero. Treat that as a vendor measurement until someone reproduces it.

Where it lands in a video pipeline

Stills come before shots. The hard problem in both is holding an identity across generations, and the 10-reference limit is the same class of feature the video vendors sell as character consistency. A character sheet, a product photo, a background plate and a style reference can go in together, which is what stops a face drifting between the still you approve and the shot you build from it.

The second use is assembly. A transparent asset drops onto generated footage inside the edit, and it survives a re-export without an alpha fringe. The third is cost. A 7B model runs on consumer cards, so iteration happens locally rather than at a per-call price, which changes how many versions of a cutout a team is willing to try.

The benchmarks are Alibaba's own

Qwen claims the model outperforms most closed-source image models, and the coverage of the release reports it as the claim it is. On Qwen-Image-Bench, a benchmark Alibaba designed and administered, Qwen-Image-2.1 scored 60.28 overall and placed seventh of 29 models, with GPT Image 2.5 Sunburst at 67.01 on the same board, according to Pondero's reading of the results. No third-party audit of this version has been published. If a figure from an earlier Qwen-Image generation is quoted next to this release, it does not apply to 2.1.

The licence is no longer Apache 2.0

Every earlier Qwen image release, including Qwen-Image 1.0, Edit, Layered and 2512, shipped under Apache 2.0 terms that place no restriction on commercial use. Qwen-Image-2.1 ships under the Qwen Research License Agreement dated September 20, 2026, which restricts use to non-commercial purposes, defined as research or evaluation, and requires a separate commercial licence from Alibaba for anyone deploying it in a product. No pricing or timeline for those terms has been published, and fine-tuned versions of the weights inherit the same restriction. A Hugging Face community thread asking for a return to Apache terms opened within hours of the release.

How to run it today

In ComfyUI, update to the latest build, download the weights from Hugging Face and drop them in your models folder, then load the Qwen-Image-2.1 template from the Templates panel and attach your references to image_1 onward. Comfy Cloud runs it without a local install. In Python, the Diffusers pipeline is QwenImage21Pipeline, installed against torch 2.4.0 or newer, transformers 5.17 or newer, diffusers from git, accelerate and pillow.

The weights are on Hugging Face, ModelScope and GitHub now, and the ComfyUI template is the quickest way in; anyone planning to ship the output inside a paid product needs a commercial agreement with Alibaba first.