EvoLink Qwen Image 3.0 (Official)

Ask most image models to render a sign, a menu, or a headline and you get decorative gibberish. Qwen Image is the family that treats text in the frame as something to actually get right - Alibaba's strong suit, and the reason its Edit line became the community's default open editor. Qwen Image 3.0 keeps that text-rendering strength as a closed API model: you can't download these particular weights, so a node like this is how it gets into your workflow.

EvoLink Qwen Image 3.0 (Official) does text-to-image and image-to-image, with a proper seed for reproducibility, a negative prompt, and a prompt-extend switch. Chinese prompts work natively - for an English speaker the model is genuinely good at rendering Chinese characters too, which is rare.

What it does

The inputs that matter:

• prompt (required) - describe the scene. If it includes text in the image, say exactly what it should read and where.
• size - auto, a ratio like 16:9, or explicit pixels like 1328x1328 (the tooltip's example - Qwen likes its native resolutions).
• quality - 1K or 2K, default 1K.
• n - 1 to 6 images per run. This is a "give me options" button; the results come back as a batch tensor.
• negative_prompt - what to avoid.
• prompt_extend - on by default in the edit node but off here; the tooltip suggests flipping it on for short prompts.
• watermark - whether to stamp a watermark. Off by default; if you're sharing drafts, leaving it off is a choice.
• seed - 0 is random, set a value to reproduce. Worth fixing once you've got a composition you like.
• image1–image4 - reference images for img2img, same 72-hour upload channel as the rest of the pack.

Outputs: image (downloaded tensor - connect Preview/Save Image or you see nothing), task_id, status_info, result_urls (24-hour expiry), response_json.

Install and key

The shared EvoLink routine. ComfyUI Manager → search EvoLink → install → full restart (close the console, don't just refresh the browser). Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/deeplearning-goethe/comfyui-evolink

Dependency is just requests; no models, no GPU. Key from evolink.ai/dashboard/keys (sk-), pasted once into api_key - saved locally, blank after, never in a shared workflow JSON.

Why you'd pick it

You'd pick Qwen Image 3.0 over a Nano Banana when the job is text: posters, packaging mockups, signage, anything where the words in the image are the deliverable. That's its lane, and it holds it well. The one thing to keep in mind is the usual API-node math - per-call cost and your prompt leaving the machine. For reproducibility, fix the seed; and if you're doing a batch of variants, remember n gives you several from one call instead of several calls. That's the difference between a sane workflow and a bill.