Nano Banana AIO

If you're going to pay Google for image generation, this is the node to reach for first. NanoBananaAIO is the flagship of the ru4ls/ComfyUI_Nano_Banana pack and the one the author's README explicitly calls "the recommended node for new workflows." It wraps Nano Banana Pro (gemini-3-pro-image-preview) - the 4K-native top of Google's image lineup, the model that made the whole "Nano Banana" name stick when it started as an anonymous HuggingFace leaderboard entry and became an official brand.

And no, there's no local model here. Like everything in this pack, it's a thin, well-built bridge between your ComfyUI graph and Google's cloud: you type a prompt, the node ships it to the API, and a finished image comes back as a normal IMAGE tensor. The trade for that convenience is a per-image bill and Google's content filters. It's one of the most-used "professional" image nodes in the community right now - Nano Banana Pro driving real Photoshop-adjacent ComfyUI workflows is a sign of where the ecosystem is leaning.

Why this node and not the NB2 one

The pack has two AIO nodes. NanoBanana2AIO runs the fast, cheaper Flash model; this one runs the Pro model with 4K native output, "thinking mode," and the sharpest image fidelity the family offers. There's also a real feature split: this node defaults use_search to on, takes up to six reference images, and keeps things simpler - no image-search toggle, no 512px option, no extreme aspect ratios. If you want the best single image the lineup can produce, this is it. If you want volume and speed, that's the other node.

How it works

Mechanically it's the cleanest node in the pack. At startup the pack reads a .env file and picks a backend: GOOGLE_API_KEY sends you down the plain Gemini API path; PROJECT_ID + LOCATION route you through Vertex AI. On execution the node converts any connected image tensors to PIL, builds a GenerateContentConfig requesting text and image output with your aspect ratio, size, and temperature, and - because use_search defaults on - attaches a Google search tool. That search grounding is the party trick: ask for "the weather forecast for Jakarta as a clean HUD graphic" and the model actually researches and cites its sources, which you can inspect in the grounding_sources output.

Set image_count above 1 and the node stops behaving like a single generator and produces a batch, running the same prompt through N sequential generations - the README's example generates four city-weather cards in one pass. That's the "all-in-one" in the name.

The inputs that matter

• prompt - everything. Multiline.
• use_search - on by default; off when you just want pure generation without web grounding.
• image_1 through image_6 - optional references for image editing or fusion. This is a 6-image node, not 14 - that's the NB2 one.
• aspect_ratio - the standard set (1:1 up to 21:9) plus Auto, added in pack v6.0.2 so the model can pick for you.
• image_size - 1K, 2K, or 4K. The Pro model's native 4K is the selling point; just be aware bigger costs more.
• image_count - 1–10, batch mode above 1.

Outputs: images (IMAGE batch), thinking (reasoning text), grounding_sources (citations). Honest fine print from the README: both string outputs only fully work on the Vertex AI path - with a bare API key they return messages nudging you to set up Vertex.

Installing

Same pack as the other nodes: ComfyUI Manager, search "ComfyUI_Nano_Banana", or:

cd ComfyUI/custom_nodes
git clone https://github.com/ru4ls/ComfyUI_Nano_Banana.git
pip install -r ComfyUI_Nano_Banana/requirements.txt

Credentials, simplest path (API key) or the more capable one (Vertex):

cd ComfyUI/custom_nodes/ComfyUI_Nano_Banana
cp .env.api.template .env        # set GOOGLE_API_KEY
# or .env.vertexai.template      # set PROJECT_ID and LOCATION

You don't strictly need google-genai --upgrade for this node (that's an NB2 requirement), but running it future-proofs you.

Gotchas

• Paid per image, and Pro is the pricey tier - the KB notes roughly $0.039–0.24 per image depending on resolution. Batch mode multiplies that.
• The filters are real. The community verdict on Nano Banana is "powerful, unimaginative, massive hidden biasing" - anything near the safety line gets refused.
• Vertex vs API matters here more than anywhere - if you care about seeing the thinking and grounding_sources outputs, skip the API key and do the Vertex setup.
• Every output carries an invisible SynthID watermark.

This is the quality pick. If you came for speed, head to NanoBanana2AIO; if you want to converse with the model and edit images iteratively, the pack's chat node (NanoBananaMultiTurnChat) runs this same Pro model.