Qwen-Image-2.0

Qwen-Image-2.0 is Alibaba's attempt to do what everyone's been threatening to do for two years: fold generation and editing into one model, and make it small enough to actually use. It's the 7B successor to the 20B Qwen-Image line - roughly a third of the size - and it unifies text-to-image and instruction-based editing in a single checkpoint. The catch, and it's a real one, is that 2.0 launched API-only in February 2026 with no open weights. The community can't download it, which means the only way to touch it is exactly this: through a cloud node like the Runware pack.

So what are you actually getting? A lighter, unified model that's fast and capable at both jobs, with one notable twist for a Qwen model: it wants big resolutions. Width and height go up to 4096, which puts it in a different league for poster-scale output than the 2048-ceiling models.

What you set

• positivePrompt (required) - plain language, generation or edit instruction depending on whether you also feed referenceImages.
• referenceImages (optional) - wire these in and the model switches into editing mode.
• width / height - 1 to 4096. Wide open, but remember: bigger is more expensive per run, and the model's sweet spot is still near the 1024-2048 range for most work.
• providerSettings.alibaba.promptExtend (default on) - LLM-based prompt rewriting. It expands and clarifies your input, and the tooltip warns it affects reproducibility: turn it off if you're chasing the exact same image from the same seed.
• negativePrompt (ignored at CFGScale ≤ 1), numberResults (up to 20), watermark, outputFormat, outputQuality.

Output is image (IMAGE).

How it works

Same pack machinery as every node: build a request with the model AIR and taskType, send over REST via the Runware SDK, get the IMAGE back. Since Qwen-Image-2.0 has no open weights anywhere, this node isn't a convenience - it's the only way most people can use the model at all. That's the whole value proposition of an API pack like this: it runs models that never left the cloud, right inside your ComfyUI graph. Cost shows on the title bar as always.

Installing

ComfyUI Manager → search Runware → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

No model downloads - there's nothing to download, which is the point. Dependencies are runware-sdk, pillow, soundfile. API key via Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Troubleshooting

Key first. Then the reproducibility gotcha: if you're setting a seed and getting different results anyway, providerSettings.alibaba.promptExtend is rewriting your prompt - flip it off. And if you're tempted by that 4096 ceiling, remember the cost scales with resolution; it's a great feature for the one hero image, a bad default for a 20-image batch.