EvoLink Nano Banana Pro (Official)

Nano Banana Pro is the top of Google's image line - Gemini 3 Pro Image, the model that made "thinking mode" a mainstream image-generation feature. It renders 4K natively, reasons about your prompt before it paints, takes up to 14 reference images, and does multilingual text in the frame better than most open models you can actually download. It's also the reason your bill can surprise you: Google prices the Pro tier per image and per resolution, and the community has genuinely hit that wall. You can't run it locally - no weights exist - so a node like this is the entire delivery mechanism.

EvoLink Nano Banana Pro (Official) is the cloud-call version: prompt and images in, finished render out, nothing running on your machine.

What it does

Text-to-image, image-to-image, and image editing from the same inputs. Plug a Load Image into image1 for img2img or an edit; your local file uploads through EvoLink's official channel and is auto-deleted after 72 hours. Up to 14 references total across image1–image4 and the image_urls_text box.

The settings that matter:

• prompt (required) - description or edit instruction, Chinese works, ~2000 tokens of room.
• quality - 1K / 2K / 4K, default 2K. 4K is where Pro earns its keep and where it costs the most. Use it for finals, not drafts.
• size - auto or ratios from 1:1 to 21:9. (Pro's ratio list tops out at 21:9 - the extreme 1:8 / 8:1 banner ratios are a Nano Banana 2 thing, not here.)
• web_search - optional web grounding for the generation. May add cost; keep it off unless you actually need current info in the image.

Note what's absent: no thinking_level knob. Pro reasons at its own pace by default - that's the "powerful but sometimes slow" character of the model. The standard outputs apply: image (downloaded tensor), task_id, status_info, result_urls (24-hour expiry), response_json.

Install and key

Same routine as every EvoLink node. ComfyUI Manager → search EvoLink → install → full restart (close the console window, not just refresh the tab). Manual:

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

Only requests for a dependency; no model downloads, no GPU required. Key from evolink.ai/dashboard/keys (sk-), pasted once into api_key - it persists locally and can stay blank after, so it never leaks into a shared workflow.

When Pro is worth it

Pro's reputation is "powerful but unimaginative" - excellent quality, less wild than it could be, and it'll follow a detailed brief precisely. That makes it a final-pass model: compose and iterate on Nano Banana 2 or the Lite tier, then send the winner through Pro at 4K. If you're reaching straight for Pro on every draft, you're paying flagship prices for workflow debugging, which is the most expensive way to debug. And the standing cloud-node caveats apply: metered per call, your images and prompts leave the machine, and moderation failures just fail (without charge). For a closed model with no local option, that's the whole trade - the trick is paying it only on the renders that matter.