Qwen-Image-2512

Qwen-Image-2512 is the big one in Alibaba's Qwen-Image family - a 20B model whose whole claim to fame is that it renders text inside images without mangling it. Logos, signage, UI mockups, multi-language captions: this is the model you reach for when you need words in the picture and they need to be spelled right. It also handles fine-grained spatial prompts better than most, which makes it a favorite for structured compositions where "the blue box on the left with 'STORE' above it" has to come out exactly like that.

You can't run the 20B stack comfortably on most consumer cards, which is precisely why the Runware pack version is appealing. No 40GB of VRAM, no fp8 shaving, no download - it runs on Runware's machines and comes back as a native IMAGE.

What you set

Required: positivePrompt, width, height (256–2048, step 16 - Qwen is pickier about dimensions than FLUX, and multiples of 16 keep it happy).

The one thing to know before anything else: Qwen runs near-distilled guidance, so CFGScale defaults off and you rarely turn it on. There's also settings.trueCFGScale for when you need real guidance - same split as FLUX, same advice: start at default.

• steps - defaults to 3. Three. That's not a typo; Qwen is guidance-distilled and wants few steps. If you set this to 30 you'll spend money for worse results.
• seedImage and maskImage - drop an image in and Qwen becomes an editor. Combined with the prompt ("change the shirt to blue"), this covers a huge chunk of image-editing work without a separate edit model.
• negativePrompt - set it, but remember it's ignored when CFGScale ≤ 1.
• numberResults - up to 20 variations per run.
• scheduler, outputFormat (JPG/PNG/WEBP), outputQuality - standard.

Output is image (IMAGE). That's it - wire it to Preview Image or Save Image.

How it works

Standard pack mechanics: the node composes a request with the model AIR (runware:... for Qwen-Image-2512) and taskType, sends it over REST through the Runware SDK, and downloads the rendered image back as a tensor. The interesting part is the LLM-encoder architecture - Qwen conditions on text the way a language model does, which is what makes the text rendering and spatial obedience work. Your node never touches the model weights; it just brokers the call and reports cost on the title bar.

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. You need a Runware API key: Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Troubleshooting

Missing key error first, always. The bigger trap is steps: people used to 20-30 step diffusion set steps high on Qwen and get muddy, over-processed results - trust the default of 3. And if you're using seedImage for edits, keep the prompt focused; Qwen edits by re-emission, so unrelated pixels can drift if you ask for too much at once.