A cut-out figure looks clean in the image editor. On a dark website, it suddenly gets a light fringe. The file really is transparent and still looks crudely cut out. This apparent contradiction is the interesting starting point for the new Qwen announcement: transparency is image information, not a finished design quality.
The original repository dates Qwen-Image 2.1 Turbo to 9 October 2026 and specifies eight denoising steps. The native RGBA output belongs to the 2.1 base, which was already described on 20 September. Release and underlying architecture are therefore separate pieces of information. Neither point on its own provides a measurement of our own of speed or image quality.
What the fourth channel changes
RGB describes colour components. The additional alpha channel describes opacity. A fully opaque area shows the colour of the subject; a fully transparent one lets the surface underneath show through. Partially transparent edges combine both. The W3C specification on compositing describes this merging of foreground and background.
For our teaching example, a drawn disc with a soft edge is enough. An edge pixel can carry some of the subject’s colour and, at the same time, low opacity. The colour you see on screen then also depends on the surface beneath it. That is not additional model intelligence but a property of how the image is later displayed.
A chequerboard preview shows where transparency occurs. It does not, however, tell you whether the edge looks convincing on your actual card colour. That is exactly why a subject can look tidy in the tool and distracting in the finished layout.
Three surfaces make a useful visual check
For a check of your own, place the same file on a light, a dark and a moderately textured surface. To begin with, change neither the size nor the image. Look at the outer edge, inner gaps and shadows separately. An unexpected light triangle between arm and body is a different weakness from a fringe running all the way round.
Don’t just record “good” or “bad”. Describe the spot: “Light fringe on the left shoulder on a dark surface.” This observation can be found again later. It says nothing yet about how often a model produces such errors. That would require several properly documented tasks and actual results.
Then scale the subject down to the intended display size. Small gaps or hard transitions can look different there than when zoomed in. A high source resolution therefore does not replace checking the image in its eventual use.
Composition remains a separate task
A cut-out subject can be moved around. That alone does not create an understandable visual narrative. Line of sight, proportions and free space for the headline are still part of the design. A figure on the right can look towards the text or away from it; the same file thereby conveys a different context.
If you want to use an externally generated image on a KENJTEC website, you add the checked file to KENJTEC MediaStudio ai manually and then select it in the relevant tool. For separately cutting out a different source image, the library offers its own function; its result is stored as an additional file under the unchanged original. This does not refer to any particular Qwen integration. Images provided in MediaStudio can be accessed via their public file address. Do not upload confidential documents there; sensitive internal files stay in the BusinessCloud, which has separate permissions.
Also note the licence: the Qwen-Image 2.1 repository is published under the Qwen Research License Agreement. It permits use only for non-commercial purposes, that is, research and evaluation; business use requires a separate commercial licence from Qwen or an offering with its own terms, which you should check first.
Eight computation steps are not yet a speed comparison
A number of denoising steps describes part of the computation. The actual run time also depends on the implementation, resolution and working environment. A comparison would have to document these conditions and the task in each case. The name Turbo does not replace this information.
The practical question is therefore more precise than “Can the model do transparency?”. It is: does the subject that matters for my purpose remain understandable on the surfaces where it will later appear, and are its transitions clean? As a first step, a small, traceable visual check is enough. The model news provides the occasion; the quality of the finished page image only becomes apparent in its actual use.
Sources and status
Research and documentation as of 25 October 2026. Qwen-Image-2.1: dated release notes for Qwen-Image-2.1-Turbo · Compositing and Blending Level 1.
