Qwen-Image-2.0-Pro

Qwen-Image is one of those names you know even if you never run it: Alibaba's 20B image model became a family legend through its Apache 2.0 licensing and the Edit line that took over open instruction-editing. Qwen-Image 2.0 was the refresh - smaller (7B), unified generation with editing - and it launched API-only in early 2026 with no open weights. This node is how you actually use it: the closed 2.0-Pro tier, rented through Runware.

That context matters because it shapes expectations. The community likes the Qwen-Image family, but the 20B base was never the daily T2I driver (smaller models won that slot), and 2.0's Pro tier is a premium API product, not a "you should switch to this" open model. Use it when you specifically want Alibaba's current image-generation quality without hosting a 20B-class model.

What you set

positivePrompt is required. Then:

• providerSettings.alibaba.promptExtend - on by default. It runs your prompt through an LLM for rewriting/expansion before generation. The tooltip flags the real gotcha: it affects reproducibility. If you want a prompt executed verbatim, switch this off; if you want better images, leave it on.
• referenceImages - IMAGE input for style/subject reference.
• width / height - 1 to 4096px, though sane values cluster around 1024; asking 4096×4096 of a 7B model is asking for artifacts.
• negativePrompt - honored only when guidance is actually applied; remember the CFGScale ≤ 1 rule from the pack.
• numberResults - up to 20, each a separate billed generation.

Output is image (IMAGE) → Preview/Save. outputFormat JPG/PNG/WEBP, outputQuality, and the optional watermark builder socket round it out.

Installing

Part of the Runware/ComfyUI-Runware pack.

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

Deps are runware-sdk, pillow, soundfile - no weights. Runware API key from runware.ai/api-keys (Settings panel or RUNWARE_API_KEY).

Where people get burned

The reproducibility trap is real and specific: promptExtend is on by default, so the same prompt+seed can still vary run to run because an LLM rewrote your words. If you're iterating on a look, turn it off for the exploration phase, then optionally back on for final quality. And keep the pricing context honest - this is the "Pro" tier of an API-only model, so it's not the frugal choice in this pack; it's the "I want Alibaba's newest and I'll pay for it" choice.