Nano Banana Pro (Google Gemini Image)

"Nano Banana Pro" is Google's flagship Gemini image model - the one that does 4K native output, chews on up to fourteen reference images, and renders text well enough that you stop noticing text is hard. This node is that model inside ComfyUI, billed through Comfy's own API gateway. You get the full Google image engine without juggling a Vertex AI project or an API key; you log into a Comfy account, top up credits, and queue.

It's a partner node, so the usual rules apply: it ships with ComfyUI core, there's no model to download, and your GPU just watches from the sidelines. What makes this one special is that it isn't only a generator - it's the model Google's whole Nano Banana branding was built around, and Pro is the "thinking" tier. Give it a messy, conversational prompt and it reasons about composition before rendering, which is why it wins the "describe the vibe, get the image" game.

How it works

You're talking to Gemini over a generate-content endpoint, not a diffusion pipeline. The model dropdown currently offers gemini-3-pro-image (the Pro engine) and Nano Banana 2 (Gemini 3.1 Flash Image, the cheaper speed tier). The node passes your prompt, optional images and files, an aspect ratio, and a target resolution to Google and returns the finished image.

Three settings matter most:

• resolution - 1K, 2K, or 4K. The 2K/4K rungs are not the model rendering at native res; they invoke Gemini's own upscaler afterward. It's good, but it's also where the cost climbs.
• aspect_ratio - auto matches your input image, or generates a 16:9-ish frame if there's no image. Otherwise pick from the usual 1:1 to 21:9 spread.
• images - up to 14 reference images (batch them with the Batch Images node). This is the editing and character-consistency path: paste a photo plus "turn her into a cyberpunk detective" and it follows.
• response_modalities - IMAGE or IMAGE+TEXT. The latter gives you a STRING output alongside the IMAGE; handy if you want the model to also caption or explain what it did.

The seed field is best-effort only. The tooltip says it plainly: fixed seeds aim for repeatable output but determinism "isn't guaranteed," and changing model or parameters changes the result anyway. Don't build your workflow around exact reproduction.

Cost and what it wires into

Pricing is token-based, and the node's price badge gives you the honest estimate per image: roughly $0.16 at 1K/2K on the Pro engine and about $0.29 at 4K; Nano Banana 2 runs about half that ($0.08–$0.18). That's the most expensive per-image node in this family, which is exactly why the badge exists. A quick "is this worth it" loop looks like: generate a look with Nano Banana Pro, then take the IMAGE and run it through your local upscaler or a Wan image-to-video node to amortize the cost across the rest of the graph.

Getting it

It's built into ComfyUI core - nothing to install. You need a Comfy account with credits (Settings → User to log in, Settings → Credits to top up), and you want a current ComfyUI since the Gemini model list rotates. Login is localhost-only unless you use a Comfy API key for LAN setups. And keep expectations calibrated on content: Google filters this heavily, and the filter is not something you can tune from the node.