Qwen 2.1 - Text to Image

Two reasons people reach for a hosted Qwen image model rather than a local checkpoint: it renders text inside images better than most local bases, and it hands back a real alpha channel when you ask for a transparent background. Both are things open local workflows still fiddle with. This node is the text-to-image half of that - prompt in, IMAGE out, no reference required.

It's also the cheapest sensible way to prototype a composition before you commit to a local render pipeline. Generate the layout, check the framing, then rebuild it locally with the model you actually want to ship.

What it's doing

A KIE Market job, model qwen2-1/text-to-image. The node submits your prompt plus the optional fields, polls the task, downloads the result and decodes it into a ComfyUI IMAGE tensor. Nothing local runs. Every execution is a billed call, and the pixels go through KIE's infrastructure - that's the trade for not owning the GPU.

Worth knowing because it changes how you prompt: enhance_prompt is on by default, so the model rewrites your words into a fuller scene description before generating. Helpful for composition, unhelpful if you're testing a specific phrasing - turn it off when the wording itself is what you're evaluating.

Inputs

• prompt - required, up to 5000 characters, any language. Be specific about the text you want rendered (wrap it in quotes in the prompt) and about the composition.
• aspect_ratio - 16:9 is the node's default even though the API's own default is 1:1; the full list covers 1:1, 4:3, 3:4, 3:2, 2:3, 9:16, 21:9, 9:21. Set it deliberately, because the pixel size is the ratio times the resolution tier.
• resolution - 1K or 2K. Only those two; 4K is rejected. 2K is four times the pixels and noticeably slower.
• background - opaque (default) or transparent. Transparent produces a genuine alpha channel in PNG or WebP. The prompt rule that trips people up: when you ask for transparency, describe only the subject. Any mention of a background, a surface or a scene tends to produce an opaque image.
• output_format - png, webp or jpeg. JPEG has no alpha, so it can't be combined with a transparent background.
• enhance_prompt - on. The single most consequential boolean here.
• seed - the reason this node is nicer to iterate on than some hosted competitors: the same seed with the same prompt and parameters reproduces the image within a service version. Lock it once you've found your composition.
• nsfw_checker - ships false, meaning no additional filtering pass beyond the provider's own.
• timeout_seconds, callback_url - defaults; leave the callback empty.

Outputs: image, url, all_urls_json, task_id, raw_json, credits_consumed, credits_left. all_urls_json is worth keeping an eye on: if the provider returns more than one result, you get every URL, not just the one the node decoded.

Install

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

Or Manager → search "KIE.ai Nodes Next". Dependencies: requests, Pillow, numpy, PyYAML. No model files, no VRAM, so the install is a few seconds. Then one key paste in Settings → KIE.ai Nodes Next → Connection - tested against KIE before it's saved, kept out of workflow JSON, shown only as a masked fingerprint.

Failure modes and honest advice

Transparency requests that come back opaque are almost always a prompt problem, not a node problem: you mentioned a background. Rewrite for the subject alone.

Also: an unwired image output means the node never executes. It's not an output node, so nothing runs and nothing is charged - a common "why is there no error and no image" moment. Connect it to a preview or save node first.

Use the right-click KIE Next · Check inputs & model controls menu when you want the documented limits without spending anything; it reads cached documentation fields locally and won't upload or submit.

And the standing objection to every node in this pack: you cannot run this model at home, so an API call is the honest price, but that price recurs on every iteration. If your actual goal is a hundred images of the same character, a local checkpoint plus a small LoRA is cheaper by the end of the week.