Nano Banana Pro API ComfyUI Setup: Nodes, Workflow and Cost

If you already run ComfyUI, adding Nano Banana Pro takes about five minutes and one login. The model ships as a Partner Node, so there is no Python environment to patch, no Google Cloud project to open and no load on your own GPU. You search for the node, wire a prompt into it, and Google's Gemini 3 Pro Image model renders up to native 4K. What takes longer is deciding where the node belongs in a graph, which settings move the price, and whether Comfy credits beat calling Google directly. This article answers all three with the numbers Google and Comfy publish as of October 2026.

💡 Short answer on cost: at Google's list price a 1K image and a 2K image cost the same ($0.134), and a 4K image costs $0.24. Draft at 2K, finish at 4K, and never choose 1K to save money, because it saves nothing.

What the Nano Banana Pro Node Does

Nano Banana Pro is the product name for Gemini 3 Pro Image, which Google exposes through the Gemini API as gemini-3-pro-image-preview. It is not a checkpoint you download. It is a hosted multimodal model, so a single node handles text-to-image, image editing and multi-image blending. The heavy lifting happens on Google's hardware while ComfyUI only sends the request and saves the result.

Gemini 3 Pro Image Under the Hood

Four traits matter when you place this model inside a graph:

• Native 4K output. The model renders at 1K, 2K or 4K with no separate upscale step.
• Up to 14 reference images. Faces, products, styles and layouts can be blended in one call.
• Text rendering in 10 languages. Posters, labels and signs come out far more legible than with most diffusion checkpoints, according to Comfy's launch post.
• Hosted inference. Nothing loads into VRAM, so a laptop with no discrete GPU runs the same graph as a workstation.

One catch: Google lists no free tier for this model on its API, so every call is billed.

Partner Node or Your Own Credential

There are three ways to reach the same model from ComfyUI, and the right one depends on who you want to send the invoice to.

Community packages exist for the third route. ComfyUI-Nano-Banana-apiyi calls Nano Banana Pro and Nano Banana 2 without a Google Cloud account, and the "Luck" Nano Banana Pro node adds timeout and retry handling, live progress, native seed modes and up to 14 stacked image inputs.

💡 A relay node sees every prompt and every reference photo you send. Treat it like any third-party service and keep client work or private photos on the official route.

Set Up ComfyUI Before You Start

Three things must be true before the node runs: a current ComfyUI build, a signed-in Comfy account and a positive credit balance. Everything else is optional.

Update and Sign In

1. Update ComfyUI. Use the desktop app updater or refresh your portable build. Older builds do not include the Nano Banana Pro node, and the official docs tell you to update first.
2. Sign in. Open Settings, then User, and log in to your Comfy account. Partner Nodes refuse to run without it.
3. Buy credits. Open Settings, then Credits. Partner Nodes call paid third-party models, so the balance has to be above zero. Monthly credits expire when the billing period ends, while top-up credits last one year from purchase.

💡 Comfy's docs also mention a permitted network environment and concurrency limits. If the node fails only on a corporate VPN, test from a plain home connection before you debug the graph.

Open the Ready-Made Template

Open the Template Library and pick the Nano Banana Pro template. Cloud users can launch the same workflow from Comfy Cloud with nothing to install. The template loads a short graph: a LoadImage node for a reference photo, the Nano Banana Pro node holding your prompt, and a SaveImage node at the end.

Upload one reference image, type a prompt, pick a thinking level and press Run. Then check your credit balance. The difference is your real cost for that resolution, and it is more trustworthy than any table in this article, including the one further down.

Build the Workflow Node by Node

The template works, but a production graph usually needs more than three nodes. Here is what each piece does and where people go wrong.

The Core Four Nodes

Settings That Change the Result

Field names shift slightly between ComfyUI versions, but current builds expose these controls:

• Prompt. Write it like a photo brief: subject, setting, light and lens. Short, concrete sentences beat long lists of tags.
• Aspect ratio. Pick the final canvas now. Changing it later means paying for a full new image.
• Resolution. 1K, 2K or 4K. This is the main price lever.
• Seed. Useful for repeating a run, but a hosted model does not promise pixel-identical output. Save the files you like.
• Thinking level. The docs list Minimal, High and Dynamic. Higher levels help with dense layouts and long text.
• Response format. Choose image only unless you want the model's text commentary too, since text output carries a small charge of its own.

Chain Multi-Reference Inputs

The 14-image limit is where this node earns its place in a graph. A repeatable pattern looks like this:

1. Feed the subject photo first, then style and prop references.
2. Say what each image is for: "Image 1 is the person, image 2 is the jacket, image 3 is the lighting reference."
3. Test combinations at 2K, then repeat the winner at 4K.
4. For two-stage work, send the first output into a second node as the only reference and ask for exactly one change.

Each extra input image costs about $0.0011 at Google's list price, so 14 references add roughly $0.015 per call. That is cheap enough to be generous, as long as every image has a job.

Draft Prompts With an LLM

ComfyUI also ships a Google Gemini node that can caption images and write prompts from rough notes, so a prompt-writing step can live inside the same graph. If you would rather draft outside ComfyUI, PicassoIA hosts Gemini 3.5 Flash for fast rewrites and Claude Sonnet 5 for longer briefs.

Ask the LLM for a prompt in four parts: subject and action, environment, lighting, camera and lens. Paste the result into the node and change one part at a time between runs, so you can tell which edit moved the image.

The Real Cost of Each Image

Official Price Per Resolution

Google's Gemini API pricing page, checked on October 6, 2026, lists these prices for Gemini 3 Pro Image:

Comfy's launch post quotes the same $0.134 and $0.24 for the Partner Node. The Partner Nodes pricing page, however, bills in credits per token: 30.38 credits per 1,000 output image tokens, 0.5064 credits per 1,000 input tokens and 3.0384 credits per 1,000 text output tokens.

A 1K or 2K image is roughly 1,120 output tokens and a 4K image roughly 2,000, which works out to about 34 credits and 61 credits before input charges. That is my arithmetic from the published table, so confirm it against your own balance after a test run.

A 4K render costs about 79% more than a 2K one. That is a fair price for print work or heavy cropping and a waste for a blog thumbnail.

Three Ways to Cut the Bill

1. Draft at 2K, not 1K. The price is identical, and you get more detail to judge composition and text.
2. Finish at 4K only for winners. Three 2K drafts plus one 4K final cost $0.642 ($0.402 plus $0.24), against $0.96 for four 4K attempts.
3. Send bulk jobs through the Batch API. It costs half as much, but results arrive on a delay, and as far as the docs show the ComfyUI Partner Node does not expose it. Bulk work belongs in a script.

Calling the API Outside ComfyUI

Sometimes the node is the wrong tool: nightly jobs, thousands of images, or an app with no canvas at all. The same model is available through Google's google-genai SDK. You need a paid-tier project in Google AI Studio, because this model has no free tier.

A Minimal Python Call

from google import genai
from google.genai import types

client = genai.Client()  # reads your AI Studio credential from the environment

response = client.models.generate_content(
    model="gemini-3-pro-image-preview",
    contents=["Overhead photo of a walnut desk with a ceramic mug, soft window light"],
    config=types.GenerateContentConfig(
        response_modalities=["TEXT", "IMAGE"],
        image_config=types.ImageConfig(aspect_ratio="16:9", image_size="2K"),
    ),
)

for part in response.parts:
    if part.inline_data is not None:
        part.as_image().save("desk.png")

Copy Prompt

Two details trip people up. image_size takes "1K", "2K" or "4K" with an uppercase K, and reference images go into contents as extra items next to the prompt. Wrap the call in a retry with backoff and log every response, because a filtered prompt returns text instead of an image.

Use ComfyUI for interactive iteration and the SDK for volume. Both reach the same model, so a prompt that works on the canvas will work in the script.

Common Mistakes and Quick Fixes

Use Nano Banana Pro on PicassoIA

ComfyUI is the right home when the model is one step in a longer graph. For a single image, a browser tab is faster. Nano Banana Pro on PicassoIA needs no nodes, no Comfy credits and no code, and the model page lists it as free to use online.

Step by Step

1. Open the Nano Banana Pro model page.
2. Write the prompt as a photo brief: subject, setting, light, lens.
3. Set the resolution to 1K, 2K (the default) or 4K.
4. Choose an aspect ratio from the 11 presets, including 16:9, 9:16, 4:5 and 21:9.
5. Add up to 14 reference images if you want a specific subject or style.
6. Pick JPG or PNG. Leave the safety filter at its default, block_only_high, unless your platform needs a stricter setting.
7. Generate, review the result, and change one thing in the prompt at a time.

Alternatives Worth Testing

No single model wins every prompt. These sit in the same text-to-image collection, so a side-by-side test costs a few minutes.

PicassoIA also offers image editing, background removal and super resolution for 2x to 4x upscaling, so a render can be fixed, cut out or enlarged without leaving the platform.

Make Your First Image Today

You now have the full picture: update ComfyUI, sign in, add credits, run the template, and keep an eye on resolution. A first test costs a few cents, and you will pick up more from one real run than from any amount of reading.

If you would rather skip the setup, open Nano Banana Pro on PicassoIA, paste a prompt from this article and press generate. Try the same brief at 2K and at 4K, then compare the results next to a second model from the table above. Picasso IA makes that comparison quick, so the next step is simple: write one prompt, render it, and see what Nano Banana Pro does with it.