Qwen-Image-3.0

Qwen-Image is the model people admire and almost nobody runs as a daily driver, because the good ones are 20B-plus and the newest ones don't ship weights at all. This node is the shortcut: Qwen-Image-3.0, Alibaba's current text-to-image flagship, served to you as a ComfyUI node over Runware's cloud. No 20B bf16 download, no 40GB card, no waiting for a quant. Type a prompt, queue, get a native IMAGE back.

It's worth knowing where this model line went, because it explains why you're reading a page about a cloud node instead of a GGUF. The Apache-2.0 Qwen-Image era ended with Qwen-Image-2.0 in early 2026, which launched API-and-chat-only. Qwen-3.x keeps that shape: the model is a hosted service, so Runware's official integration is effectively the only way it shows up inside ComfyUI. The pack's job is to make that feel local - the node looks and wires like any other generator, and the request runs on Runware's backend instead of your VRAM.

Mechanically these Runware nodes are thin. Each one is generated from a schema and talks to the Runware API over the runware-sdk; when you queue, your prompt and any reference images are uploaded, inference happens server-side, and the result comes back as a standard ComfyUI image tensor. That's the whole trick, and it's why installing it pulls almost nothing heavy: runware-sdk, pillow, soundfile - no model files at all.

The inputs that matter are few. positivePrompt is the only required one, and it's an LLM-encoded model, so prompt it in full sentences rather than comma-tag soup. referenceImages takes a normal ComfyUI image and uses it as a subject or style reference. There's a real negativePrompt, which is rarer on modern models - Alibaba kept it, and it's respected. width and height default to 1024, and seed plus numberResults behave like you'd expect from any sampler. Two settings worth touching: providerSettings.alibaba.promptExtend defaults to on and silently rewrites your prompt through an LLM to "expand and clarify," which the tooltip warns affects reproducibility - flip it off when you're chasing an exact result - and safety.checkContent adds a moderation pass that slows every run. The advancedFeatures.watermark.* group lets you stamp a text watermark into the output if your use case needs it; ignore it otherwise. The output is image, which you wire straight into Preview Image or Save Image.

Install and the API key

The pack is one install for all 20-plus model nodes, so this is the same for every Runware node:

• ComfyUI Manager: search Runware, install, restart.
• Or manually: cd ComfyUI/custom_nodes && git clone https://github.com/Runware/ComfyUI-Runware, then pip install -r ComfyUI-Runware/requirements.txt, restart.

You then need a Runware API key from the dashboard. Paste it in ComfyUI Settings → Runware API key, or set the RUNWARE_API_KEY environment variable, or run runware auth login once from a terminal. Missing key is the classic first failure - the node errors out before it ever queues.

Gotchas worth knowing

This is a pay-per-call node. Each run shows its cost on the node's title bar (something like $0.00078 · NSFW: no), and it adds up faster for image batches than you'd guess. Your prompt and reference images also leave your machine and go through Runware's moderation, which is the standing tradeoff of every cloud node - fine for Qwen-Image-3.0 because there is no local version to compare against. And because it's cloud inference, "it's slow" usually means queue or network, not a wrong setting; if a run fails, check the API key first, then the console for the actual error message the SDK prints.