Analysis September 24, 2026 workflow 9 min read

The official ComfyUI templates for Qwen-Image-2.1 can produce genuine transparent cutouts with no selection tool, but the sentence that switches transparency on is hidden inside a note in the workflow file and is missing from the documentation page most people will read instead.

Qwen-Image-2.1 ComfyUI image-gen local-creative-ai creative-workflows prompt-craft licensing-provenance

Qwen-Image-2.1 Will Hand You a Transparent PNG, But Only If You Say the Magic Sentence

The official ComfyUI templates cut the background out inside the model. The instruction that makes it happen is buried in a sticky note, and the docs page never mentions it.

There is an instruction you have to type to get a transparent image out of Qwen-Image-2.1, and it does not appear on the official documentation page for the workflow. It appears on a sticky note inside the workflow file itself, which you only read if you open the JSON in a text editor or click a note node most people scroll past.

Here it is, copied exactly from that note:

This is an RGBA format image with transparency. [your description]. The image has an alpha channel and a transparent background.

You replace the middle. That's the whole trick. And it matters more than a hidden tooltip should, because what comes back is not a photo with the background painted white. It is a picture with genuine transparency, the see-through kind a PNG can hold and a layout program will respect, produced by the model rather than by a selection you made afterwards.

Anyone who has spent an afternoon feathering edges around hair will understand immediately why that is interesting. Anyone who has not can think of it this way: normally you generate a picture and then cut the subject out of it, which is a second job with its own mistakes. Here the subject arrives already cut.

Why the hidden sentence is worth an article

Qwen-Image-2.1 is an open-weight image model from Alibaba's Qwen team, released this month, and its ComfyUI packaging picked up official workflow links yesterday morning, September 23. ComfyUI is the free node-based tool a lot of people use to run image models on their own computer. "Open-weight" here means the model files are downloadable and sit on your drive rather than behind an API, which is a nice property to have in a week when two hosted video models got retirement dates.

The transparency is the part of this release that changes actual work. Everything else it does well, the 2K output, the sharp small text, the reference-driven editing, it does well alongside other models. The see-through output is a different category. The official ComfyUI documentation describes the mechanism plainly: the part of the model that turns its internal picture back into pixels carries four channels instead of three, so the fourth one, the transparency, comes out of the model directly instead of being reconstructed later.

So this is one of the rare capabilities that deletes a step rather than speeding one up. For anyone making icons, stickers, product shots for a white page, logo marks, or game sprites, that step was the tedious one.

And it is one sentence away from working, and that sentence is not where you would look for it. I went through the official docs page for these templates line by line. It covers sampler settings, resolution targets, how the reference-image slots are numbered, how the caching node behaves. It does not tell you how to ask for transparency at all.

The two routes, and they are different jobs

What the docs page does give you is a third template I had not seen mentioned anywhere else, and between them the two routes cover different situations.

Route one: generate a cutout from nothing. Use the text-to-image template with the magic sentence. Good for assets that do not exist yet: an icon set, an illustrated object, a stylised product.

Route two: remove the background from a photo you already have. There is an official template called Remove Background that reuses the editing setup with a fixed instruction, and the instruction is refreshingly unglamorous. The docs quote it directly: Remove the background, and output a PNG image. The workflow then shows your result beside the original so you can see what it took away.

That second route is the one I would reach for on client work, because it operates on the photograph they sent rather than on something invented. It also means the same model file serves both jobs, no swapping.

Put this into practice

Total time if your ComfyUI is already up to date: about fifteen minutes, most of it downloading. The files are large.

1. Update ComfyUI first. The official guide is blunt about this: if you cannot find these templates in the template library, your install is too old. Do this before downloading anything, because a missing node in a loaded workflow is almost always this.

2. Download three files. From the Comfy-Org Qwen-Image-2.1 repository, take the int8 versions, which just means the compressed ones that need less memory and are what the templates load by default:

ComfyUI/models/diffusion_models/  qwen_image_2.1_int8_convrot.safetensors
ComfyUI/models/text_encoders/     qwen3vl_8b_int8_convrot.safetensors
ComfyUI/models/vae/               qwen_image_2.1_vae_bf16.safetensors

There are full-precision versions of the first two if you have memory to spare. Start with the smaller ones. You can always swap later, and the difference will not be what decides whether the transparency works.

3. Load the text-to-image template. Search "Qwen-Image-2.1" in the template library, or grab image_qwen_image_2_1_t2i.json directly. The top level is only five boxes and two of them are notes, because the real machinery is collapsed inside a single group. That makes it unusually easy to read for a ComfyUI graph.

4. Type your prompt inside the frame. The full sentence again:

This is an RGBA format image with transparency. [your description]. The image
has an alpha channel and a transparent background.

Do not paraphrase it on the first try. Get one working result with the exact wording, then experiment.

5. Leave the save format as PNG. The template uses a save node whose PNG path handles the transparency. Change the format to JPEG and the see-through parts turn into a colour, usually a bad one, and you will blame the model.

6. For full-size output, set the resolution target to about 4 megapixels. The template defaults to 1 megapixel, roughly 1024 by 1024. The docs say the model generates natively at 2K, so 4 megapixels gets you 2048 by 2048. The workflow's own note adds a second rule of thumb the docs page leaves out: prefer dimensions in multiples of 32.

7. For the background-removal route, load image_qwen_image_2_1_background_removal.json instead, drop your photo into the image-loading box, and leave the instruction alone. The default already says the right thing.

Where this breaks

I want to be specific here, because three of these will cost you a result and look like the model's fault.

The template runs at roughly half the step count its own note recommends. Steps are how many passes the model makes before it hands you the picture; more passes, more time, more detail. The docs confirm all three templates ship at 25 steps. The note inside the graph says the official Qwen pipeline uses about 40 to 50 with the sampler these templates use. Nobody explains the gap. On solid shapes you will not notice. On hair, smoke, fur, and lace, which are exactly the edges you wanted the model to cut for you, you probably will. Push it to 40 and see whether your subject needs it.

Your negative prompt does nothing. There is a dial in the graph labelled cfg, which sets how hard the model is pushed to obey your wording, and these templates set it to 1. The note states plainly that the negative prompt does nothing while it sits there. So you can type a long list of things you do not want, get no error, and get no effect either. The note says raise cfg only if you actually intend to use a negative prompt. This is the single most confusing failure mode in the whole setup, because silence reads as success.

On the editing and background-removal routes, the canvas size can shift your picture. The docs warn that when you turn on the custom-size option you should keep it close to the size of the resized reference image, otherwise the edit can drift. Cheapest safe setting: leave the sizing alone on your first few runs.

Two of the magic sentences are in circulation and they are not identical. The wording I quoted is from the note inside the workflow file, verbatim. The model card's own suggested phrasing is close but not the same, differing in a few small words. Neither source says whether the difference matters. I would not assume it does, and I would also not go hunting for the perfect incantation. Get one working result, then vary one thing at a time.

Three files in the repository have nowhere to plug in yet. The repository also contains two prompt-enhancer variants with pe_t2i and pe_i2i in their names and a smaller four-bit reader, all uploaded on September 20. The repository README does list them and does say what they are for, prompt enhancement through a separate text-generation node. The docs page never mentions them, and nothing in the three shipped templates is wired to use them. So you can download them, put them in the right folder, and still have no graph that touches them. I mention them because they point at something useful, which is that the part of the model reading your sentence is now a swappable piece with several sizes, and that is where a lot of the memory goes. The plumbing to take advantage of it is not in the templates.

And the licence is the real constraint. The packaged files are published under qwen-research, per the model card, pointing at Qwen's own licence terms. Read them before a cutout you made this way goes into a paid job. This is the part that is easy to skip when a capability is this satisfying, and it is the part that matters most if you are billing for the result.

What I'd try next

The thing I keep turning over is the second route, background removal by written instruction. Every background remover I have used is a specialist: a model trained on one job, with one dial, that fails in predictable ways on glass and hair. This one is a general image model being told what to do in a sentence, which means in principle you can say something more interesting than "remove the background." Remove the background but keep the cast shadow. Remove everything except the two people on the left. I do not know whether any of that works, because the template ships with one fixed instruction and the docs do not encourage wandering off.

That seems like the most promising hour anyone reading this could spend, and it costs nothing but electricity. If you try it and find a phrasing that reliably keeps a shadow, I would genuinely like to see it, because that is the one thing a specialist cutout tool has never been able to do for me.

Medium metadata

Title: Qwen-Image-2.1 Will Hand You a Transparent PNG, But Only If You Say the Magic Sentence
Subtitle: The official ComfyUI templates cut the background out inside the model. The instruction that makes it happen is buried in a sticky note, and the docs page never mentions it.
Tags: ComfyUI, Image Generation, Qwen, Design Workflow, Open Source AI
Primary sources: ComfyUI's Qwen-Image-2.1 workflow guide · Comfy-Org/Qwen-Image-2.1 on Hugging Face · image_qwen_image_2_1_t2i.json · image_qwen_image_2_1_background_removal.json