Gemini 3 Pro

You know what Nano Banana Pro is, even if you've never said the name out loud - it's Google's flagship image model, the thing the community calls "powerful but unimaginative" and yet keeps going back to for its 4K native output and genuinely good text rendering. This node is a bridge: it drops gemini-3-pro-image-preview (the API name for Gemini 3 Pro Image) straight into a ComfyUI workflow, with your images and prompt going out and a proper IMAGE tensor coming back.

First, kill one expectation: nothing runs locally. No weights, no VRAM, no diffusion model lurking behind it. You hand this node an image, it phones Google's API, and Google does all the work. That means you need a Gemini API key and an internet connection, and every generation costs money (roughly $0.04 to $0.24 per image depending on resolution). In return you get the model's party tricks - native 4K output, real text-in-image accuracy, and multi-image editing - without leaving ComfyUI.

What it does under the hood

The node converts your image tensor to PNG bytes, bundles it with your prompt, and streams the request through the official google-genai SDK. Response comes back as chunks; it grabs the inline image bytes, turns them into a standard ComfyUI image tensor, and collects any accompanying text along the way.

The genuinely interesting part is the resilience layer. You can feed it up to three API keys, and if key #1 hits a rate limit or errors, it silently moves to key #2, then #3. If all three fail, it waits 15 seconds and loops the whole sequence - up to 10 times - before giving up. That's aimed squarely at overnight batch renders where a single 429 would otherwise kill your run at 3 a.m.

The inputs that matter

• api_key_1 - your primary Google Gemini API key. This is the one you actually need; api_key_2 and api_key_3 are optional backups that only matter if you're running long batches against free-tier quota.
• prompt - your text instruction, a multiline field. The default is "Make this image cyberpunk style," which is fine for a first test.
• resolution - 1K, 2K, or 4K. This is native output, so 4K is a real 4096×4096, not an upscale. Watch the price tag climb with it.
• image_1 / image_2 / image_3 - up to three reference images to edit against. You can leave them all disconnected for pure text-to-image.
• model_name - locked to gemini-3-pro-image-preview, one choice. Don't go looking for a dropdown that isn't there.

Outputs are processed_image (a normal IMAGE - wire it into Save Image, Preview Image, or keep processing it) and text_response (a STRING with whatever the model said alongside the image, handy as a prompt for downstream nodes or just to read what it thought it did).

Installing it

Easiest path: ComfyUI Manager → Install Custom Nodes → search ComfyUI-Gemini3 and let it handle the rest. Or the manual route:

cd ComfyUI/custom_nodes/
git clone https://github.com/Daohoi123/ComfyUI-Gemini3.git
cd ComfyUI-Gemini3
pip install -r requirements.txt   # google-genai, pillow, numpy

Then restart ComfyUI and look under Gemini AI → Gemini 3 Pro. On Windows with the embedded Python, swap pip for ..\..\..\python_embeded\python.exe -m pip. The one heavyweight dependency is google-genai, and there are no model files to download - the model lives on Google's side.

Where people get burned

The biggest gotcha isn't in this node at all - it's the model. Nano Banana Pro is heavily safety-filtered, and that shows up here in a confusing way: if the API's filters kill the image, the node raises "Model returned text only (No Image)." A clean prompt, a fresh key, and a moment of patience usually fix it; a celebrity or copyrighted-IP prompt may simply never work.

Other things to expect:

• ImportError: No module named 'google' - you skipped the pip step. Run the install above.
• 404 / "model not found" - this is a preview model. Your key needs access to gemini-3-pro-image-preview in Google AI Studio; not every key has it by default.
• 429s on free tier - normal, and exactly what the multi-key retry loop is for. Two or three keys genuinely help overnight.
• It's a small, quiet pack. Zero community footprint, one node, no active issues to speak of - you're trusting the README and the MIT license here. The source is short enough that you can read the whole thing before you rely on it.

Honest verdict: for a one-off Google edit, ImageFX or AI Studio is faster and free. This node earns its keep the moment you want Gemini output feeding the rest of your graph - a sketch restyled to 4K, text dropped into a render, then back into your local upscaling and compositing stack. If that's your workflow, it's a clean little wrapper, and the overnight-batch retry logic is more thought than most API nodes bother with.