Qwen Image 2.1 · Edit

Two things make this the version of Qwen editing worth defaulting to. It takes up to ten reference images instead of the three its 2512 and 1.0 siblings allow - which turns "swap this jacket" into "here is the person, the room, the product, the two other angles and the colourway". And it can hand back a genuine alpha channel, so a cutout exits the node as data rather than as a black-and-white picture you have to guess at.

Price: from about 5.4 credits ($0.054) per image. The "from about" is because each reference you add raises the quote.

What comes out, exactly

Three outputs: images, alpha, and task id. The alpha is a ComfyUI MASK where 1 is opaque - same polarity as the mask feeding a compositor, not inverted. That's the useful half of the pair for downstream work: multiply, composite, or drive a Load Image-style cutout without ever thresholding an RGB render.

One thing to know before you wire it up: the returned RGB is pre-multiplied by alpha. Transparent regions come out black, not checkerboard-transparent, because the API returns a PNG that remembers the colours hiding behind the transparency and the node deliberately zeroes them so the cutout reads correctly. So the alpha output is where the transparency actually lives. Save both.

The fields that matter

prompt - "Describe the subject, composition, lighting, and any wording to render." This model does typography well; if a sign or label is part of the scene, ask for the actual words.

image_urls - an Autogrow group. One socket to start, another each time you connect something, up to ten. Ordered, so connect them in the order you'll refer to them. A batch on one socket counts as several pictures.

mask_url - optional, single IMAGE, and the one input people don't expect: a black-and-white mask for local editing where white is repainted and black is kept. Use it with exactly one reference image. If you've ever wished instruction editors respected a boundary, this is that.

background - opaque by default. Set transparent to get the subject alone on alpha - and note the tooltip's warning: it needs output_format set to PNG or WebP, so a JPEG at transparent background gives you nothing.

enable_prompt_expansion - on by default, which surprises people. The service rewrites your prompt into a fuller scene first. Usually it helps; when you're chasing an exact composition it's the reason the model keeps adding things you didn't ask for. Turn it off for word-for-word adherence.

resolution - 1k tier by default, with 2K available on this family.

aspect_ratio - 1:1 default. On an edit that means your source gets reframed, so check this before blaming the model.

seed - fixed by default; repeated queues reuse the cached result instead of charging twice.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Restart, or install through Manager (search PopNow). No dependencies are installed - the pack declares none beyond what ComfyUI ships - but it needs Python 3.10+ and the V3 node API, so a very old ComfyUI won't load it. Key from popnow.ai/api/console/keys into Settings → PopNow, then restart again to pull your account's catalog.

Practical gripes

Qwen-Image-Edit re-emits the whole frame. Ten references gives it more to average over, and the standard failure is drift on the parts you said nothing about. Say "keep" explicitly, or use mask_url and make it true.

Transparent background plus multiple images in one batch: pictures of different sizes get resized to the first one's dimensions, because they have to share a tensor. Don't mix a portrait cutout and a landscape cutout in one run and expect both intact.

And the usual money notes: PopNow 40201 is an empty balance, 40004 is an unservable combination - it names the field to change - and a cancelled run keeps going on PopNow's side and is charged if it succeeds. Set Settings → PopNow → Max credits per run before you start adding references.