Qwen 2.1 - Text to Image

What it is

KIE's hosted Qwen text-to-image endpoint, wrapped as one ComfyUI node. Prompt in, image out, no model files, no VRAM, per-call billing.

The honest read on the model family: the Qwen-Image 20B base is respected locally but it's not what people reach for as a daily driver - it's heavy, and smaller models like Z-Image outrun it on consumer cards (see kg/entities/qwen-image.json). What Qwen's line is genuinely famous for is editing - Qwen-Image-Edit is the open instruction editor everyone built on. This node is the generation side of that family on someone else's GPU.

The distinguishing feature here isn't quality, it's background: transparent. Getting a real alpha channel out of a hosted generator, on demand, is the thing I'd actually wire this into a workflow for: product cutouts, sprite work, anything you'd otherwise spend a segmentation pass on.

How it works

The node is built from KIE's request schema for the endpoint. Your prompt and settings go out as a task, the node polls until the job is done (default timeout_seconds of 1200, i.e. 20 minutes), then the returned image is downloaded and handed back as a native ComfyUI IMAGE. If a widget is left alone, its documented default is what gets sent.

Because the node is generated from the live catalog, the widgets here are exactly the fields KIE documents - nothing invented, nothing hidden.

Inputs and outputs

Required:

• prompt (STRING) - describes the image. Any language, up to 5000 characters; longer input returns a 422.

The optional settings that matter:

• background - opaque or transparent. Transparent gives you a real alpha channel and requires PNG or WebP; it can't be combined with JPEG. When you use it, describe only the subject. Prompts that mention a background, scene or surface tend to produce a non-transparent result.
• resolution - 1K or 2K. 1K runs roughly 7–12 seconds, 2K roughly 30–55 seconds. Anything else (including 4K) is a 422.
• aspect_ratio - nine options from 16:9 to 9:21. Watch this one: the widget ships defaulted to 16:9 while the underlying doc text says 1:1, so set it explicitly rather than trusting the default.
• enhance_prompt - on by default. It rewrites your prompt into a fuller scene description first, which per the author "noticeably improves quality". Leave it on unless you're testing prompt phrasing.
• seed - same seed plus same prompt and settings reproduces an image within the same service version.
• nsfw_checker - defaults to false, and the tooltip explains the semantics: off means KIE's content filtering is disabled and results come back from the model as-is. Be honest with yourself about what that does and doesn't guarantee.
• output_format (png / webp / jpeg), timeout_seconds, callback_url.

Outputs: image (the picture, wire it to a preview or a Qwen image-to-image node for a second pass), url, all_urls_json, task_id, raw_json, credits_consumed, credits_left.

Install

Manager: search KIE-Nodes-Next, install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next

The pack's requirements.txt is small - requests, Pillow, numpy, PyYAML - so installation is quick. Windows users can use the one-click .vbs installer from dist/.

Then save your key: Settings → KIE.ai Nodes Next → Connection → Paste / replace KIE.ai API key. It's validated before it's stored, and you get a masked fingerprint and a ✅ CONNECTED status with your credit balance. KIE_API_KEY as an environment variable still works and takes priority, which is what you want on a headless box.

Where people get burned

Two defaults disagree, so decide for yourself. That 16:9-versus-1:1 aspect ratio mismatch is the kind of thing that produces "why is my square prompt coming out wide" posts. Pick the ratio you want.

Transparency is a naming thing, not a slider. If your prompt says "on a marble countertop", you asked for a background and you'll get one. Describe the object; let the alpha do the rest.

Filtering off is not the same as unfiltered. The tooltip is candid: filtering can't be guaranteed to catch everything, and results are returned as-is. That cuts both ways.

Credits are reported, not capped. There's no budget guard in the node. If you're looping prompts, watch credits_left - or wire it to a text node and log a run.