Nano Banana Pro (own API key)

Nano Banana Pro is Google's flagship image model, and in ComfyUI there's only one way to run it: over the API, because there are no weights to download. The stock Nano Banana Pro (Google Gemini Image) node does call the model - but it routes through api.comfy.org and bills Comfy credits, with no field for your own Google key. This node is the "wait, I already have a Google account" move: it talks to Google directly with your key, on your own billing, and skips Comfy's metered storefront entirely. If you've been burning prepaid credits on a model you could be paying Google for directly, that's the whole pitch.

The pack is fresh off the shelf - MIT, a single author, no community track record yet - so give it a skim before the first run. The code is short, readable HTTP against Google's API; nothing hiding in it.

How it works

Pick a backend on the node, because that decides everything:

• AI Studio (api key) - posts to generativelanguage.googleapis.com with your x-goog-api-key. Simple, but Google has excluded Gemini API from the $300 Cloud free trial since March 2026, so you're paying per image with prepaid credits.
• Vertex AI (service account) - posts to aiplatform.googleapis.com with an OAuth token from your gcloud login or a service account. Same model, but billed to your Cloud project, which is where the $300 trial credit applies. The pack's reason to exist.

Behind the scenes it sends your IMAGE tensor as PNG base64 inline_data to :generateContent and turns the returned image back into a tensor the graph can save. Nice touch: Vertex serves the same model under a different id (gemini-3-pro-image vs the AI Studio gemini-3-pro-image-preview), and the node translates that so you never see it.

The inputs that matter

Most of the widget list mirrors the built-in node. The ones you'll actually touch:

• model - gemini-3-pro-image-preview is the flagship default; drop to gemini-2.5-flash-image for the cheaper, faster tier.
• aspect_ratio and resolution - ratios from auto through 21:9, output at 1K/2K/4K. 4K is the Pro model's party trick, and where the price jumps.
• seed - capped at int32 because that's what Google accepts, with control_after_generate for iteration.
• response_modalities - IMAGE+TEXT also returns the model's written reply on the text output; IMAGE skips it.
• images - leave unconnected for pure text-to-image, or feed up to 14 reference images (stack them with a batch node) for image-to-image work.
• system_prompt - pre-filled with a "your only output is a picture" guard so the model draws instead of explaining. Leave it until the model starts refusing.

The default timeout of 300s is generous for 4K; bump it only if big jobs keep dying mid-request.

reference_mode - stop it copying faces

This is the interesting bit, because Gemini has no reference-strength slider. Attach a photo of a person and the model tends to drag their face into the result no matter what your prompt says. The node steers with a trailing instruction appended after the images - where image models weight wording hardest. Default prompt first (no face copy) tells it to borrow mood, lighting and palette but invent a person who matches your words; match subject (copy likeness) preserves the face instead. It's instruction steering, not a hard constraint, and it only runs when images is connected - but in practice it's the difference between "looks like him" and "the woman you asked for."

Install

cd ComfyUI/custom_nodes
git clone https://github.com/haristahir1/comfyui-gemini-ownkey

or search the pack name in ComfyUI Manager, then restart. The AI Studio backend needs nothing extra - just requests, torch, numpy and Pillow, which ComfyUI already ships. Vertex needs one package on Windows:

python_embeded\python.exe -m pip install google-auth

Put your key in config.json (or the api_key widget / GEMINI_API_KEY env var) and check it free with python_embeded\python.exe ComfyUI\custom_nodes\comfyui-gemini-ownkey\check_key.py --list-only. Vertex setup is the fiddly part: run both gcloud auth login and gcloud auth application-default login - the second is the one Python reads and it's easy to miss - then enable aiplatform.googleapis.com on your project.

Where people get burned

Most failures are billing and permissions, and each has a telltale error. 429 prepayment credits depleted means your AI Studio balance hit zero - top up or switch to Vertex. 403 SERVICE_DISABLED means you skipped the gcloud services enable step. A 404 on Vertex usually means your project doesn't serve that model at the global location - try "location": "us-central1" in config.json. And "Vertex mode needs credentials" is the missing application-default login. One security habit worth keeping: prefer config.json over the widget, because widget values get saved into workflow JSON and error reports - a key typed into the node travels with every workflow you share. On Vertex no key is sent at all, so leave api_key blank there.