Qwen-Image-Edit

Qwen-Image-Edit is where a huge class of image work stopped needing masks. It's the open instruction editor from Alibaba's Qwen team: you hand it a picture and a sentence - "remove the person in the background," "swap the jacket to leather," "make it golden hour" - and the change just happens. Apache 2.0, uncensored by policy, and good enough that by 2026 it had absorbed the jobs that used to take a mask plus a ControlNet plus an IP-Adapter. Its own LoRA library is the secret weapon: pose, angle, next-scene, and expression LoRAs trained specifically on the editor, and they don't work on the base Qwen-Image model at all.

The honest limit, and its own users say it first: the model re-emits the whole frame, so pixels nobody asked about can drift, and faces degrade across a chain of edits. The standard fix is bolting a mask back on around it. But for a single clean instruction edit, nothing in the open world is easier.

What you set

• referenceImages (required) - the image you're editing, wired in as an IMAGE. One or more; more refs means more context for consistency.
• positivePrompt (required) - the instruction, in natural language. This is not a tags box; write "turn the car red" as a sentence.
• width / height (128–2048, step 16) - the output canvas; keep it near the source image's resolution.
• steps - defaults to 8. Qwen-Edit wants far fewer steps than SD-era editing; don't crank it.
• CFGScale - off by default; the editor runs near-distilled guidance. settings.trueCFGScale is there if you genuinely need more adherence.
• lora socket - this is where Qwen-Edit's LoRA library plugs in. Pose LoRA to re-pose a character, next-scene LoRA for sequence continuation, etc.
• negativePrompt (ignored at CFGScale ≤ 1), seed, scheduler, numberResults (up to 20), outputFormat, outputQuality.

Output is image (IMAGE).

How it works

Like every node in the Runware pack: it builds a request with the model AIR and taskType, sends it over REST through the Runware SDK, and downloads the result as a native IMAGE. The edit happens entirely in the cloud on Qwen's hardware - no 20B model download, no local VAE dance. The node prints cost on the title bar; edit runs are short, so it's cheap to iterate. One thing to remember: because Qwen re-emits the whole frame, each edit is a fresh generation - chain edits one at a time and check between them rather than piling all your changes into one prompt.

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; dependencies are runware-sdk, pillow, soundfile. API key via Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Troubleshooting

Missing key error first, as always. If edits look "floaty" or unrelated pixels shift, that's the re-emission behavior - keep instructions surgical, or add a mask in a workflow that crops the region. If nothing changes at all, check referenceImages is actually wired (it's required for a reason). And if you need heavy structural changes across many steps, that's what the Edit-Plus node is for.