Qwen - Text to Image

Which Qwen are we doing today

The menu here is genuinely confusing, so here's the plain version. The pack exposes two generations of Qwen image endpoints: the older qwen/* set (this node, plus Qwen - Image to Image and Qwen - Image Edit) and the newer qwen3/* set (Qwen3 Text to Image, Qwen3 Pro ...). They don't share a parameter vocabulary. The Qwen3 group talks in resolution tiers and aspect ratios; this group talks in scheduler-era terms - steps, guidance scale, acceleration, and a fixed list of named sizes.

So which do you want? Honestly, the Qwen3 ones are the newer models and I'd default there. This node is worth reaching for when you want the knobs: a specific seed you can actually hold onto, a step count you control, and a negative prompt that behaves the way you expect. It's also the Qwen endpoint most tutorials and screenshots reference, which matters when you're following along.

Inputs worth your attention

• prompt (required, up to 5000 characters) - nothing else is mandatory. Nice for a quick connectivity test: paste anything, press run, see if your key works.
• image_size - square_hd by default, with square, portrait_4_3, portrait_16_9, landscape_4_3, landscape_16_9 alongside it. Named sizes, not free dimensions, so "make it 1216×832" isn't an option here.
• num_inference_steps - default 30, up to 250. More steps cost more time, not necessarily more credits.
• guidance_scale - default 2.5 with a 0–20 range and 0.1 steps. On these Qwen endpoints low guidance is normal; don't port your SDXL 7.0 habit over.
• negative_prompt - default is a single space. That's a hint about how much it matters here.
• acceleration - none / regular / high; the docs recommend high for images without text, which is a nice way of admitting it hurts typography.
• seed, output_format, enable_safety_checker (on by default), nsfw_checker (off), timeout_seconds, callback_url.

Outputs: image, url, all_urls_json, task_id, raw_json, credits_consumed, credits_left. The credit pair is the pack's signature move - after each run the node itself tells you what those credits bought, so you can price a batch before you commit to one.

Install

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

Restart ComfyUI afterwards. The pack's Python dependencies are light (requests, Pillow, numpy, PyYAML), and the README's heavier infrastructure - a Windows one-click installer, a full uninstaller, an offline bootstrap catalog - lives in dist/ if you'd rather not touch a terminal on Windows. Then set the key once at Settings → KIE.ai Nodes Next → Connection. It's validated before the old one is replaced, and it never lands in workflow JSON, which matters if you share graphs.

Things that catch people out

The widget set is generated from KIE's live documentation at startup. That's excellent for keeping up with new models and slightly strange the first time a node you already use grows a new dropdown. Old workflows still run - new widgets come with their own defaults - but don't be surprised if the panel doesn't look like the screenshot you followed.

Also, enable_safety_checker and nsfw_checker here are two different fields with two different defaults, and it's easy to flip the wrong one and conclude the provider is filtering you when it's your own setting. Read the labels. Whichever way you set them, the filter that matters runs on KIE's side and on Alibaba's model before that; a hosted endpoint is not a censorship workaround, it's just someone else's policy layer.