Qwen3 Pro Text to Image

What it is, and who it's for

One text box, one IMAGE output, and none of your GPU involved. Qwen3 Pro Text to Image wraps KIE.ai's qwen3/pro-text-to-image endpoint, so a prompt typed in ComfyUI turns into a hosted generation and comes back as a normal image tensor you can feed straight into a local upscaler, a mask, whatever.

That's the whole pitch, and it's worth being blunt about when that's useful. If you own a decent card, a local checkpoint is free after electricity and this isn't. If you don't - laptop, integrated graphics, a Mac, a work machine you can't install a 20 GB checkpoint on - this node is one of the cheapest ways to get a competent modern generator into a real graph, priced per call instead of per month. The RB community's read on resellers like KIE is consistent: cheap, one key, no subscription lock, and your inputs leave the machine, which is the deal.

How it works

The pack builds one ComfyUI node per callable page in KIE's API documentation, at startup, from the docs themselves. That's why this node has a very specific hand-picked set of widgets rather than a generic "prompt + JSON" form: the fields come from the endpoint's documented parameters.

On run, the node posts a generation task and polls it until the result arrives. timeout_seconds defaults to 1200 seconds (20 minutes), adjustable from 30 to 7200. Once it completes you get the image downloaded into your ComfyUI environment, plus the KIE-side URL kept as a string.

The inputs you'll actually touch

• prompt - the only required field. Qwen3 handles Chinese and English; the placeholder default is a sunset cityscape if you want to see it work immediately.
• prompt_extend - default on. The author's tooltip is explicit that when enabled the model optimises your positive prompt and "significantly improves results for simple descriptions." It's a prompt-rewriting step you can't see, which is either a feature or an annoyance depending on how much you care about reproducibility. For a shot list where every frame must match a mental image, turn it off and write properly.
• resolution (1K/2K) and image_size (1:1, 3:2, 2:3, 4:3, 3:4, 16:9, 9:16, 21:9) - these are the money dials. 2K costs more than 1K and that's per image, every reroll included.
• negative_prompt, seed, nsfw_checker (off by default), output_format, and a callback_url you should leave empty for normal use.

Outputs are the pack's standard seven: image, url, all_urls_json, task_id, raw_json, credits_consumed, credits_left. The last two are the honest part - the node prints what the call cost and what's left on the node itself, so you find out you burned your balance at three in the afternoon rather than at the end of the week.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
pip install -r ComfyUI-KIE-Nodes-Next/requirements.txt

Four dependencies, all light. Restart, then set the key once: Settings → KIE.ai Nodes Next → Connection → Paste / replace KIE.ai API key. Validating happens before the old key is overwritten, and it's stored in ComfyUI's user directory, not in the workflow. On a headless box, KIE_API_KEY in the environment takes priority.

Traps worth knowing

Seeds are not a reproducibility guarantee on a hosted model - the endpoint can be updated server-side without telling you, so treat a seed as "stable-ish" rather than a lock. Model IDs also come and go on KIE; if this node starts erroring with a "not resolved" message, refresh the catalog rather than debugging your graph. And the 4K end of the range is where a batch of twenty suddenly costs real money - do your exploration at 1K.