🚀 VRGDG NanoBanana Pro 🚀

VRGDG_NanoBananaPro calls Google's Gemini image-generation API and returns the result as a ComfyUI image. The name is the product: Nano Banana Pro is Google's flagship image model (Gemini 3 Pro Image), and this node is a thin API client for it - paste in an API key, type a prompt, get an image.

Let's get the two most important facts on the table. First, it needs a Google API key (api_key input, from the Gemini API console) and it's a paid API - every generation costs real money. Second, this is a closed, heavily filtered model: the KB's coverage of closed-source models flags Google's image line as "aggressively" censored and, on aesthetics, "powerful but unimaginative." You are renting a very capable but opinionated renderer, not running something local.

Inputs

• api_key - your Google API key. This is the thing people forget, and without it the node just fails.
• prompt - the generation prompt. Default "A cinematic wide landscape."
• model - the model picker: gemini-3-pro-image-preview (the flagship, default) or gemini-3.1-flash-image-preview (the speed/quality hybrid tier).
• image1 … image4 - up to four reference images. This is the secret weapon: feed it character stills or scene frames and the model generates variations that stay anchored to them, which is exactly what a music-video pipeline wants for consistent look-development.

Output is a single image, ready to drop into the rest of your graph - VAE encode it for an I2V pass, save it as a reference, or feed it to an upscaler.

Why you'd reach for it

Two genuinely good uses. First, look development: before committing to a video run, use the reference images to iterate on a character's look at 4K-native quality with Google's "thinking mode" reasoning, then feed the winner downstream as your LTX I2V anchor. That's the KB-documented workaround for character drift in LTX - generate consistent stills first, animate them second. Second, upscaling/refining: Comfy Org's own upscaling handbook names Nano Banana Pro as a current-generation upscaler, so the node doubles as a very capable one-shot image refinement step. The text-in-image and multi-language rendering are also genuinely strong, if those matter to your workflow.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl

Restart ComfyUI; it's under VRGDG/NanoBananaPro. ComfyUI Manager: search vrgamedev. No model downloads - the dependency is the API key and the network.

Honest note

This node is a convenience wrapper around an API you could call yourself, and the wrapper adds exactly the ComfyUI-native image handling. Budget for it: generations cost per image and add up fast in a batch workflow, and every output carries Google's invisible SynthID watermark. And temper expectations on the aesthetic side - the model is technically spectacular and stylistically constrained. For keyframes and reference material in a music-video pipeline, that trade is often worth it; for creative freedom, a local model will fight you less.