Qwen Image 3.0 · Edit

Every other Qwen image node in this pack tells you, in its own tooltip, that there is no negative prompt field and that you should write exclusions into the main prompt instead. This one has the field. That's the tell that 3.0 is a different generation of the family, and it's also the most practical change for anyone doing edit work: "no extra people, no text on the wall" now goes where it belongs.

About 4.5 credits ($0.045) an image, with readable signage and headlines as the headline capability. Typography is the family's whole reputation and 3.0 is where it's sharpest.

Fields

prompt - describe the target image: subject, composition, lighting, and any text you want rendered. Refer to your references in the order you wire them. Quoted words get rendered as words, which is the trick for signage: a shopfront with a hand-painted sign reading "OPEN LATE" behaves very differently to a shopfront sign.

negative_prompt - the new one. Describe what to exclude. Keep it short and concrete; long negative lists on modern models mostly add noise. Two or three specific things beats a comma-separated wall of "blurry, lowres, worst quality".

image_urls - Autogrow, ordered, one socket growing per connection. Reference images in the order your instruction uses them.

resolution - 2K default. An actual image-resolution tier, unlike the LoRA variants of 2.1 which cap at 1k.

num_outputs - the schema says minimum 1, maximum 1. It exists as a widget and can't be changed. One prompt, one image, one charge - don't go looking for a batch.

aspect_ratio - 1:1 by default and it will reframe your source. Set it deliberately; on an edit, reframing is the most common way to lose the picture you liked.

enable_prompt_expansion - True by default. The service rewrites your brief into something fuller before generating. Turn it off when you're chasing an exact composition, or when you're rendering a specific string of text and don't want it rephrased.

seed - fixed, so re-queueing the same graph reuses the cached image rather than buying another one.

Outputs: images and task id. No alpha channel on 3.0 - if you need transparency, that's Qwen Image 2.1 · Edit.

Is it worth it over 2.1?

For type, yes. 3.0's whole pitch is legible signage and headlines rendered into the frame, and the difference against older Qwen revisions is the difference between a sign that reads and a sign that suggests reading. For general editing - wardrobe, relighting, cleanup - 2.1 at a similar price with a transparency option and a mask input is the more flexible tool.

The other honest caveat is the one that applies to the family: editing re-emits the whole frame, so drift is real. Chain two or three edits and faces wander. One edit then composite.

Install

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

Restart, or Manager → Custom Nodes Manager, search PopNow, Install. There is no dependency to manage: the pack's metadata lists none, and its HTTP, image and video work all leans on libraries ComfyUI already ships. Python 3.10+ and the V3 node API.

Then popnow.ai/api/console/keys → Settings → PopNow → paste → restart once more. Yes, twice. The node catalog is built at startup from the models your key can reach, so anything new only appears on the next boot. On a server, set POPNOW_API_KEY instead of using the settings field - a key in a widget is a key in every workflow you share.

Codes worth knowing by heart

No PopNow API key is set / PopNow 401 - key problem, not a model problem. 40201 - no credits. 400 naming a field - that field got a value the model rejects; with this node it's usually resolution or aspect_ratio. 40004 - the combination isn't servable, and the message names the setting to move. 40310 - verify your email. content_rejected / generation_failed - refused or broken, refunded.

Set Settings → PopNow → Max credits per run to something like 45 and a ten-render exploration can't surprise you. Each node shows the going rate on its badge and quotes the run before anything is held.