Nano Banana Pro

"Nano Banana" was the community's nickname for Google's Gemini image model, and it stuck because the thing was absurdly good at instruction-following - you could tell it to add a banana to a scene and it actually would. Nano Banana Pro (google:4@2) is the bigger tier of that line, and the remarkable part of this node is what you don't need: no Google AI account, no separate Gemini API key, no Vertex AI billing page. One Runware key and Google's model just shows up in your node menu.

This is a closed model by definition - you were never going to run Gemini locally - so the cloud wrapper isn't a compromise, it's the only way in. What you get for the API call is the model's real superpower: reference-image editing. Wire images into referenceImages and you can do character-consistent generations, style transfers, and multi-image composition the way Google demos them. That's the headline.

How it works

Standard Runware imageInference with the google:4@2 AIR. Your prompt and reference images go up as base64, Google renders, and the result comes back as a native IMAGE tensor. Two Google-specific knobs are worth flagging before you touch anything else:

• providerSettings.google.webSearch - live web-search grounding. Turn this on and the model can incorporate real, current facts ("the current skyline of Osaka" instead of a hallucinated guess). It's the most Google-y feature on the node and the one nobody else offers.
• providerSettings.google.safetyTolerance - high/medium/low/none, with off meaning "use Google's defaults." Google's image models are cautious; this is the dial if a prompt keeps getting refused.

The inputs that matter

• positivePrompt (required) - and it's worth being specific. This model follows instructions unusually well.
• referenceImages - the IMAGE input that unlocks editing, subject consistency, and style transfer. The single most important optional input on the node.
• size - a dropdown of 34 presets, from "from input" (match the reference) up through 1K/2K/4K at every common ratio. One control replaces the width/height pair.
• settings.systemPrompt - a global instruction that shapes every generation.
• numberResults - up to 20 variations, which is a lot of different seeds for one credit.
• settings.temperature / settings.topP - gated on/off toggles with paired value fields; off uses the model default.

Installing it

Install the pack, not the model. 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

Key from runware.ai/api-keys → ComfyUI Settings → Runware API key (or RUNWARE_API_KEY). Nothing downloads.

Where people get burned

• Safety is the model, not the node. Google's filters are the strongest in the field. If generations get refused, safetyTolerance only goes so far - the model's guardrails are load-bearing.
• Web search costs and surprises. webSearch is a paid, slower mode that changes results every time (it's live data). Use it deliberately, not by accident.
• Reference quality is ceiling quality. Editing works best with clean, high-res inputs; a noisy reference drags the output down.
• It's metered. 20 results on one run is 20 billed generations. The title bar shows the total; check it before you batch.

Nano Banana Pro is the node you reach for when you want Google-grade editing without Google's onboarding. It's the closest thing in the pack to a magic wand - just mind the bill.