Qwen-Image-2.1 Makes Transparent PNGs Without a Cutout Step. Then the Licence Shows Up.
An open image model that generates the hole in the background instead of making you cut it, what you can actually make with it today, and the licence line that decides whether any of it can leave your hard drive.
Ask almost any image generator for a sticker and you get a sticker sitting on a white square. Then comes the part nobody writes about: the selection tool, the refine-edge dialogue, the background-removal service that costs money and still leaves a pale halo around loose hair. That chore has survived four years of image models getting better at everything else.
Qwen-Image-2.1, which went public on Sunday, hands back a file that already has a hole in it.
It is not the first thing to do this. LayerDiffuse, published in 2024, bolted transparency onto Stable Diffusion and got packaged into ComfyUI and Forge, and a few commercial tools offer a transparent-background switch. What is different here is that the transparency is part of the base model rather than an attachment, and the weights are downloadable.
Not a white square you then remove. An actual transparency channel, generated alongside the colours, in the same pass. The piece of the model that turns its internal picture back into pixels was rebuilt with a fourth channel for opacity, so "how see-through is this pixel" is something the model decides while it draws, not something you subtract afterwards. That is why the edges behave on hair, on glass, on smoke, on all the places where a selection tool falls apart.
There is a catch large enough that I nearly did not write this up at all, so let me put it on the table early: the licence says non-commercial. We will get to exactly what that means. First, what the thing does.
What you can make with it today
The obvious use is sticker sheets and logo marks, and it is very good at those. The less obvious uses are where it starts saving real hours.
Product cutouts for a catalogue. Overlay elements for a video edit, where a halo around the edge is the difference between a comp that sits in the shot and one that floats above it. Interface icons that drop onto an unknown background. Smoke, sparks, dust and light-leak elements, which are the classic nightmare because they are semi-transparent all the way through and a hard selection destroys them. A model that generates partial opacity natively is drawing exactly what a compositor would want handed to them.
The other capability worth knowing is reference handling. You can pass it up to ten photographs in one go, and the published examples are the ones you would actually want: a group portrait assembled from six separate headshots, a full outfit built from five images: a model, clothing, shoes, a bag and a hat. Not a style transfer. Subjects carried across, faces held.
And it edits by being pointed at, rather than by being masked. Draw a rough circle on the input image, or just scribble over the region, write one plain instruction per circle, run it once. The reference example changes three things in a single pass: takes off a watch, changes hair colour, replaces an item of clothing. Three circles, three sentences, one run. The reason that works is that the part of the model that reads your prompt also reads images through the same path, so a circle drawn on the picture is an instruction in the same sense that a word is.
Text rendering is good, which matters for anything with a sign or a label in it. More on that shortly, because somebody measured it properly and the way they did it is worth stealing.
Default output is 2048 by 2048, and the team publishes a table of recommended sizes for each aspect ratio. Use those. Typing your own dimensions is the fastest way to get a worse picture out of any model of this kind.
Try it in the next ten minutes
The lowest-friction route needs nothing installed.
1. Open the free demo. huggingface.co/spaces/Qwen/Qwen-Image-2.1. It runs in a browser tab. There is a queue at peak hours.
2. Write the prompt in the shape the model expects. This is the part that trips everyone up, and it is not optional. The transparency is triggered by a specific phrasing the model was trained on, and it wraps your description on both sides:
This is an RGBA image with transparency. A cute cartoon dragon sticker. The image has alpha channel and the background is transparent.
Swap "a cute cartoon dragon sticker" for whatever you want. Keep the opening sentence and the closing sentence exactly as they are. Drop either one and you will get a perfectly nice opaque square and wonder what went wrong. I would treat those two sentences as a template you paste every time rather than something you rewrite.
3. Save as PNG. JPEG has no transparency channel, so a JPEG download has already lost it.
4. Check it against a coloured background. Drop the result onto mid-grey or a saturated colour in whatever you edit in. Transparency errors are invisible against white and obvious against anything else. Fifteen seconds, and it saves you finding out later.
That is the whole first loop. If it works for you, the next two steps up:
Running it on your own machine. ComfyUI supported it from day one and ships two ready-made setups you can drag straight into the canvas: a text-to-image template and an editing template. The matching model files live at Comfy-Org/Qwen-Image-2.1. Budget an evening and a large download.
On a Mac. Somebody published an Apple Silicon conversion on Monday. On an M5 MacBook Pro with 32 GB of memory it finishes a 1024 by 1024 image in 3.7 to 4.2 minutes, roughly two and a half times faster than the standard route. The download is just under 15 GB. The buried finding in that write-up is worth passing on to anyone who owns a Mac: running this on the Neural Engine is the wrong choice. Measured on part of the model, the graphics processor beat the Neural Engine by a factor of seven.
The measurement I want more people to copy
Most compression write-ups report how far the smaller version drifts from the full-size one, and call that quality. That test has a hole in it: it defines the full-size version as perfect, so the smaller one can only ever lose.
The team at ModelsLab who published four-bit builds of this model on Monday ran a different test on their Blackwell-card build. Eight random seeds, two prompts with lots of text in them, and then they sat down and read all 48 resulting images by eye, scoring whether the sign said what the prompt asked for. The full-size model got 14 out of 16. The compressed one got 15 out of 16.
And then they refused the headline. Their own words, where "bf16" is shorthand for the full-size model: "A single prompt where the quantized build beats bf16 is trajectory variance, not evidence of superiority." They also tried an automatic text-reading pass first, found it was scoring random background text rather than the sign, and threw it out. Their other build, the one that runs on a far wider range of cards, was never put through this test, and they say so rather than letting the good number cover both.
If you ever publish a comparison, that is the standard. Score against what you asked for, not against the bigger model.
Where it breaks
The licence is the real limit, and it is not a footnote. Qwen-Image-2.1 ships under the Qwen Research License Agreement. Section 2(a) grants use "FOR NON-COMMERCIAL PURPOSES ONLY," and the agreement defines non-commercial as "for research or evaluation purposes only." Selling the output requires a separate licence from Hangzhou Tongyi Laboratory, which you would have to go and ask for. This is not ambiguous and it is not a grey area you can reason your way around by saying the output is yours.
So: a sticker pack on a storefront is out. Assets in a client deliverable are out. A personal project, a test, a look at whether this workflow is worth building around, all fine. The community conversions inherit the same terms, and the honest ones say so on the front of their model card, which is more than plenty of repackagers manage.
Your graphics card will fill up faster than the file sizes suggest. The picture-drawing half is about 10 GB at full size and compresses to under 5. The part that reads your prompt does not compress along with it, and in the measured setup it accounted for roughly 16 GB of a 21.5 GB total. Which means the four-bit version does not solve your memory problem, because the memory problem was never the file everyone is compressing. That deserves its own article and it is getting one.
The prompt template is load-bearing and badly signposted. The phrasing is out there, since every repackaged copy reproduces the model card, but nothing in a typical interface tells you that two specific sentences control whether you get transparency. You find out by getting opaque squares.
The Mac conversion is a narrow slice. Text-to-image only, one fixed size, no editing, no reference images, prompts capped at 64 tokens, and the prompt-reading model is a separate errand. Read the limitations section before starting a 15 GB download.
Faster is a claim about your specific card. The same four-bit build that runs 1.84 times faster on a 4090 runs at 0.35 times the speed on an RTX PRO 6000 Max-Q. Slower. The authors printed that number rather than hiding it, which is how you should read every speed claim: as a measurement on one machine, not a property of the file.
What I would actually do with this
Test it this week on the specific thing that annoys you most. For me that is semi-transparent elements, because thirty years of selection tools have never solved smoke, and a model that draws partial opacity directly is the first approach that attacks the problem from the right side.
Then keep the results in a folder marked as a test, because that is what they legally are.
The open question is whether native transparency becomes a standard feature or stays attached to restrictive licences. Nothing in the approach is exotic. Any lab could add a fourth channel to the piece that turns pictures back into pixels. Whoever does it next under permissive terms will get a great deal of attention very quickly, and I suspect that is months away rather than years.
If you find a prompt shape that gets cleaner edges on hair than the stock template does, I would like to see it. That is the part I have not cracked.
Medium metadata
Suggested tags: AI Image Generation, Design Tools, Open Source AI, Graphic Design, Machine Learning
Subtitle: An open image model that generates the hole in the background instead of making you cut it, what you can actually make with it today, and the licence line that decides whether any of it can leave your hard drive.
Primary sources: Qwen-Image-2.1 model card · GitHub repository and release note · Qwen Research License Agreement · ModelsLab four-bit build and measurements · Core ML conversion for Apple Silicon · ComfyUI model files