Qwen Image supports transparent assets you can keep editing

On September 20, 2026, Qwen released Qwen-Image-2.1. It brings text-to-image generation and image editing into one model, supporting images with transparency, edits to transparent layers and extraction of subjects from photographs.

The practical difference is that the result can be a separate visual asset. A cup without a background, for example, could sit on several practice layouts with different colors. You could then ask to change its handle color. The model provides this generation and editing capability; you still need to inspect the shape, edges and transparent areas it produces.

You do not have to decide the whole layout in the first generation

Suppose you are practicing a book-club poster. You want to arrange the title, time and colors yourself, and only need a cup illustration. A finished image with its own background brings that scene into your layout. A separate transparent asset lets you try a light background, then a dark one, and judge whether the subject remains clear.

This separates what to draw from where it belongs. First assess the cup’s shape and proportions, then its relationship to the text. If only the handle color is unsuitable, try changing that detail and compare the rest of the image instead of starting over with an entirely new description.

Editable results still need inspection

A checkerboard in a preview does not by itself prove transparency. Put the result on both light and dark backgrounds. Look for white fringes around the cup, check whether the hole in the handle reveals the background, and make sure parts of the subject have not disappeared. If an edit also changes the rim, proportions or shadow, that local change has not yet met your requirements. The poster and cup are suggested evaluation tasks, not completed tests.

Use the project’s official online demo; you do not need to install the model locally first. Check the conditions before starting: the research license limits its noncommercial scope to research or evaluation, and commercial use of the model materials requires a separate license. Access to a demo does not imply unrestricted commercial use. The small task below is for evaluating the capability.

Follow the demo’s current usage instructions:

1. Generate from text first. Leave the input-image area empty. Describe both the subject and transparency, for example: “Create a simple illustration of a blue ceramic cup, showing the complete cup and handle, with no text or logos. Output an RGBA image with an alpha channel, transparent everywhere outside the cup.” This is a suggested test prompt; substitute your own object.
2. Choose whether to enhance the prompt. Prompt enhancement is enabled by default. To evaluate your original wording directly, turn it off, then select Generate Image. If the page reports that the demo is unavailable, keep your prompt and wait for it to return; a loading page is not a generated result.
3. Inspect the transparent asset. The result uses PNG format. Keep the original PNG with support for transparency, then place it in your layout tool on light and dark backgrounds. Check edges, holes and the subject itself. A screenshot of the preview is not a substitute for the asset file.
4. Change one detail, then compare. Put the generated image into the input-image area to enter editing mode. Try: “Change only the cup handle to green. Keep the cup’s shape, all other colors and the transparent background.” Compare the new result beside the previous one. Changes to unrelated parts mean it still needs adjustment.

A complex poster is unnecessary for a first evaluation. Start by checking whether you can obtain a simple asset with acceptable edges that you can continue editing, then decide whether this kind of tool belongs in a longer production process.