Making a single banner can take half a day just for cutting out the assets: tracing the outlines of people or products, redoing the shadows and blending them into the background. It is time closer to manual labor than to design work.
On September 20, 2026, Alibaba's Qwen team released the image generation and editing model "Qwen-Image-2.1" as open weights. One of its features is that it can output RGBA images, which carry transparency in addition to color information, directly from text instructions. This could make the cutout step itself unnecessary. Whether a production company can use it in deliverables, however, is decided by the license before performance.
Transparency is not the only thing it can do
According to the distributor, text-to-image generation and image editing are unified in a 7B-parameter generator. The edits it lists include changing a subject's expression while keeping the background transparent, rewriting text inside an image, and combining up to 10 reference images into a single composition.
What matters in production is this part: being able to edit while keeping transparency. If you generate an image with a background and then cut it out, flaws tend to appear around hair and semi-transparent areas. If the image carries transparency information from the start, the step of fixing those flaws may be reduced.
Look at the license category first
The model is released under the Qwen Research License Agreement (dated September 20, 2026). The original license text grants use only for non-commercial purposes, that is, research or evaluation purposes, and states that using it for commercial purposes requires obtaining a separate commercial license from the distributor. The original text also gives the contact for applying. In other words, evaluating it internally and creating images for a client's website are treated differently, even when the operations are the same.
"Released as open weights" and "free to use commercially" are two different things. Even for a model whose weights are distributed and can be run locally, the distributor sets the license terms. In production work for clients, creating assets without checking this turns into a question of who bears the cost if replacements are needed after delivery.
Check the following three things, in this order.
- The model's license category — whether commercial use is allowed or it is limited to research or evaluation purposes. If limited, where to apply and on what conditions
- How your own generated output is handled — whether records show who generated what in which environment. Whether this also covers work by subcontractors
- Contract terms with the client — whether they permit the use of AI in deliverables. Whether there is a disclosure obligation
If you overlook the third, problems remain even if the model allows commercial use. How clients define the scope of AI use in contracts is covered in AI usage terms and acceptance for deliverables.
Also check what it takes to run it locally
Another thing that is easy to overlook is the environment it runs in. 7B is the parameter count of the part that generates images; reading the instructions uses a separate Qwen3-VL 8B. The weight files you have to handle are not small. In the weights distributed for the image generation tool ComfyUI, even the lightweight format (INT8) comes to 7.26GB for the generation part and 9.35GB for the part that reads the instructions.
It will not necessarily run as is on production laptops, so decide first whether to put it on a shared workstation or distribute it to individual machines. If you leave this undecided and only pass around the message that it seems usable, everyone tries it in their own environment and the origin of the resulting assets becomes unclear.
One more factor in deciding between in-house and outsourced production
Once the license conditions are in place and you can produce transparent assets in-house, there is less need to get a quote every time for a simple banner swap. On the other hand, for production assets that require brand consistency, having a person decide the composition and colors is likely to keep rework down better than using generated output as is.
This dividing line will not shift much even as models change. The criteria for the decision are laid out in whether to produce images with generative AI in-house or outsource them.
If you test it, do so somewhere separate from production
For the evaluation itself, it is safer to set up one internal dummy project and test with that rather than with materials from a real project. While the license is limited to research or evaluation purposes, keep a separate working environment on the premise that outputs are not reused in production.
What to look at then is the downstream work rather than generation accuracy: whether the output PNGs drop straight into your production tools, whether semi-transparent edges remain as intended, and whether the same instructions produce a similar image again. If these three do not pass, the cutout time may disappear, but it will just be replaced by time spent on fixes.
We checked the original license text (Qwen Research License Agreement) in the distributor's GitHub repository and the descriptions in Qwen's official blog and repository on September 24, 2026. The weight file sizes are as shown on the distribution page for ComfyUI. We have not run the model, checked the quality of its output or applied for a commercial license. This article does not offer an interpretation of the license, and the distributor's original text takes precedence for its terms. Always check the latest original license text before deciding on commercial use.
For setting up asset production for web development or creating rules for generative AI use, please consult GleamHub.
