Qwen Image 3.0 Pro · Text to Image

What this node is for

Qwen-Image has one reputation that outlived everything else about it: it renders words. Signage, packaging, posters, Chinese characters - commercial-grade typography was the headline claim when the 20B weights landed in August 2025, and it's the reason people still reach for the family over Flux. The catch is that 20B parameters of anything is a bad fit for a 12GB card, and the line went API-first - Qwen-Image-2.0 arrived in February 2026 with no downloadable weights at all. Qwen Image 3.0 Pro is the current top of that stack, and this node is how you get it into a ComfyUI graph: Alibaba's model runs on PopNow's servers, and you get an IMAGE tensor back you can wire into a mask, an upscaler, or Save Image.

So the honest reason to reach for it isn't "I want a Qwen generator." It's "the text in this image has to be correct" or "I don't want to pull 40GB of weights and argue about samplers." Pro here is the detail tier - lifelike skin, fabric, glass, per the node's own summary - and it costs roughly double the standard 3.0 at 2k.

How it works

The pack ships one node per PopNow model, and each node's widgets are generated from that model's published input schema. Nothing is sampled locally. When you hit queue, the node uploads any connected media, builds a request from your widget values, quotes the job, submits it, polls until it finishes, then downloads the result and hands it back as a normal ComfyUI value. Prices are quoted before the hold, and a failed run is refunded - which matters, because the second half of every cloud-node decision is what happens when it doesn't work.

Two things the pack does that are worth knowing: your API key is never a node setting (it lives in user/popnow/config.json or POPNOW_API_KEY), and the seed starts on fixed. Queue the same graph twice and you get the same picture back rather than a surprise second charge.

The inputs that matter

prompt is where the effort goes. Qwen's text encoder is an LLM, so it reads sentences, not tag soup - describe the subject, the composition, the lighting, and write the exact text you want rendered inside quoted punctuation. The tooltip tells you to include any text to render; do that literally, because this model is the reason you're here.

enable_prompt_expansion defaults to True, and that's the trap. The service rewrites your brief before generating, which helps vague prompts and quietly sabotages careful ones - your hand-typed headline is now "improved." For typography work, turn it off. If you leave it on, prompt_optimization_mode decides how hard the rewrite goes: direct is a single pass, agent is a multi-step agent rewrite for when your prompt is a one-liner and you want the service to do the thinking.

resolution (1k / 2k) and aspect_ratio (1:1 through 21:9) are your framing and cost levers; 2k is the expensive tier. negative_prompt takes what to exclude, and num_outputs gets you more than one roll per run. Outputs are images - a batch of IMAGE tensors - and task id, a string worth routing to a Note node when you're debugging, since it's what PopNow's API console and support use to find the run.

Installing it

# ComfyUI Manager: search "PopNow", Install, restart.
# or:
cd ComfyUI/custom_nodes
git clone https://github.com/popnow-ai/popnow-comfyui

Then create a key at popnow.ai/api/console/keys, restart, paste it under Settings > PopNow, and restart once more - the node list is rebuilt from your account's catalog at startup. On a headless box, POPNOW_API_KEY=sk-popnow-... skips the settings dance. There are no model downloads and no Python dependencies beyond what ComfyUI already ships. If you used the old single-file node, delete custom_nodes/popnow_api.py before installing the pack.

Where it goes wrong

PopNow 40201 means the wallet is empty - buy credits or drop to 1k. PopNow 40004 means that combination of settings isn't servable; the message names the field. An error ending in request_id ... is the one to paste to support. And if a headline comes back garbled, check expansion first, then whether you gave the model too much text for the resolution - 2k is not just more pixels, it's where small type survives.