Qwen-Image

Qwen-Image is Alibaba's open-weights image model - Apache 2.0, no non-commercial rider - and the thing that made it famous in the ComfyUI world is that it's an editing model, not just a generator. Feed it an input image and a prompt and it will restyle, modify, and composite the way FLUX's Kontext used to be the only open model that could. This node (runware:108@1) runs it on Runware's cloud, so you get the model's strengths without a single checkpoint download or a VRAM budget.

The "no local run required" part matters more here than it usually does. Qwen-Image locally is an ecosystem project - fine-tunes, ControlNet variants, the works - but all of that is a setup project. This node is the shortcut: install the Runware pack (ComfyUI Manager → Runware, or git clone https://github.com/Runware/ComfyUI-Runware into custom_nodes/), add an API key in ComfyUI Settings, and Qwen-Image is a node you can queue from a fresh install.

How it works

Standard cloud image inference under the hood, with a twist in the tuning: Qwen-Image is guidance-distilled. The node's CFGScale_value defaults to 1, and the negative prompt is ignored when guidance is disabled (CFG ≤ 1) - the tooltip says it outright. This is the 2026 model-class behavior that still trips up anyone with SDXL muscle memory: guidance is baked into the weights, so cranking CFG doesn't buy adherence, it overcooks. The editing power comes from the seedImage socket - wire in an image, prompt the change, and the model rewrites it.

Inputs that matter

• positivePrompt (required) - descriptive prose; Qwen-Image's LLM-encoded prompt path understands sentences, and that's what unlocks the editing behavior.
• seedImage - the editing input. This is the input that makes Qwen-Image Qwen-Image.
• width / height - 256–2048. Unlike some SD-derived models, it isn't locked to 1024 multiples.
• maskImage / outpaint - for inpainting and extending; maskMargin adds context pixels around the masked region for cleaner blends.
• steps - direct INT, default 20 (no toggle - Qwen-Image exposes it plainly).
• CFGScale - toggle-gated; the CFGScale_value default of 1 is the guidance-distilled sweet spot. settings.trueCFGScale exists for true CFG if you insist.
• lora / acceleratorOptions / watermark / ultralytics - builder sockets.

The image output is a native IMAGE into Preview or Save, with per-run cost on the title bar.

Where people get burned

• CFG 1 is not broken, it's the design. Raising CFG on Qwen-Image past a small range burns the image and roughly doubles render time (that's what the unconditional pass costs). Leave the gate off.
• Negative prompts do nothing at CFG ≤ 1. There's nothing for the negative to attach to when guidance is off. If you need exclusion, that's not this model's tool.
• Editing needs a good seedImage. Qwen-Image edits what it can see; a low-res or muddy input image caps how good the result gets.
• The toggle pattern - CFGScale, strength, maskMargin are "enable to set X" booleans with paired _value widgets. Flip the toggle, set the value.

If you've been eyeing Qwen-Image for its editing chops but the local setup felt like a project, this is the zero-effort way in. It won't replace a fine-tuned local stack, but it's the model's core behavior with none of the maintenance.