Qwen Image 2 · Edit

Qwen Image 2 is the odd one in the family, and the oddness is the point. Qwen-Image-2.0 folded generation and editing into a single 7B model - a real architectural change from the 20B family everything else here is built on - and then shipped it API-only, with no open weights. There is no checkpoint to download, no GGUF, no Civitai page. A hosted node is the only door to this model that exists.

So this node is what you get: a prompt, one photo, an aspect ratio, a seed. 5.25 credits ($0.0525) per image.

The trade you're making

Compared to the Qwen Image 2.1 edit two doors down, this node has fewer controls and costs more. No resolution tier, no prompt rewriter, no background/alpha option, no output format choice, no mask, no LoRAs. What it has instead is a newer, smaller, unified model that does the instruction-following part briskly and without ceremony.

Whether that's worth the premium is a judgement call, and mine is: reach for 2.1 unless you specifically like 2's rendering. Less to tune is sometimes exactly what you want; it is not automatically better.

The three things you set

prompt - the tooltip is edit-shaped: "Describe the image to produce from the source: subject, composition, style, and lighting." Note it says describe the image, not the change. This model wants the finished picture described, with the source as the starting condition.

image_urls - an Autogrow group whose tooltip says "Exactly one". It's the plural field name with a single-image job, which is a wrinkle of the schema rather than a suggestion, so connect one socket and leave it. There's a sentence in the tooltip about only processing images you're authorised to use, which is the API being explicit that your photo leaves the machine.

aspect_ratio - and here's the trap. Default is 1:1, and the tooltip spells it out: "It does not follow the source image." So a wide photo edited on default settings comes back square and recomposed. If the shape of your source matters, set the ratio to match it before you queue. This is the difference between "an edit" and "a new picture that shares a subject with my old picture", and it surprises people every time.

seed - fixed by default, so an accidental re-queue returns the cached result instead of a second charge.

Outputs: images, and task id.

How the node exists at all

Nothing here is hand-written per model. The pack fetches PopNow's catalog and every model's published input schema at ComfyUI startup and generates one node class per model-and-task, which is why an API-only model like this appears as a node with exactly the settings the API takes - and why a model nobody wrote a node for shows up after a restart once PopNow publishes it. Your photo is uploaded to PopNow's storage as a private reference, the job is quoted and submitted, and the result lands back as an ordinary IMAGE you can wire anywhere.

Install

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

Restart. Manager → Custom Nodes Manager → PopNow → Install does the same with a mouse. There's no dependency step: zero declared dependencies, everything comes from the aiohttp, torch, Pillow and NumPy ComfyUI ships. Python 3.10+ and the V3 node API are the requirements.

Then a key from popnow.ai/api/console/keys into Settings → PopNow, and restart once more so the model list comes from your account. Working on a remote box? POPNOW_API_KEY in the environment, so it never lands in a workflow you share.

Failure modes, honestly

Recomposed output: your aspect ratio. Wrong subject or ignored instruction: your prompt describes a change rather than a result. PopNow 401 is a bad key, 40201 an empty balance, 400 names a rejected field, 40004 an unservable combination, 40310 an unverified email. Refused generations come back content_rejected and are refunded.

Set Settings → PopNow → Max credits per run before you experiment - 5.25 credits a render adds up at the pace this node queues.