Replicate google/nano-banana-2

Nano Banana 2 is Google's speed-quality hybrid in the Banana lineup - under the hood, Gemini 3.1 Flash for images, released February 2026. It's the middle child: not the 4K flagship (that's Nano Banana Pro) and not the original consumer product. What it does that the others can't is reach out to the internet while it generates. This node from ComfyUI-API-DockerCPU drops that whole thing into your ComfyUI graph as a plain node, with no GPU and no local checkpoint required.

Where this one earns its keep is the search grounding. Tell it to draw something time-sensitive - "the current F1 standings podium," "today's weather over the harbor" - and with google_search on, it pulls real-time web info and renders from that. Turn on image_search and it grabs web images as visual context, which is a legitimate way to do reference-based generation without curating your own image set. It's a genuinely novel capability for a ComfyUI node, and it's also exactly the kind of thing that sounds cool until you remember it's Google deciding what's relevant.

How it works

Like every node in this pack, it's generated from the model's Replicate schema. Prompt goes out as a prediction against google/nano-banana-2, and back comes an IMAGE tensor plus API_JSON (the exact payload sent, for debugging). Closed model, hosted inference, zero local compute.

The inputs that matter:

• prompt - as always with Google's image models, natural language wins. This one is a Gemini Flash-class model, so it's snappier than Pro but a rung below it on raw quality.
• google_search / image_search - the headline feature. Both default off; flip them on when the prompt depends on current information or external imagery.
• IMAGE_1 through IMAGE_14 - up to 14 reference images, same multi-image editing and character-consistency story as Pro.
• aspect_ratio - this one goes wider than Pro: 1:4, 1:8, 4:1, 8:1 and friends. That's panorama territory, and it's a real differentiator if you're building banner or background assets.
• resolution - 1K default, up to 4K. Lower resolution is faster and cheaper; the schema's own note says higher takes longer.

output_format is jpg or png. The node's other outputs: IMAGE and API_JSON, same as every node in the pack.

Install

ComfyUI Manager (search "ComfyUI-API-DockerCPU") or:

cd ComfyUI/custom_nodes
git clone https://github.com/trustypangolin/ComfyUI-API-DockerCPU
cd ComfyUI-API-DockerCPU && pip install -r requirements.txt
export REPLICATE_API_TOKEN="your_token_here"

Restart ComfyUI and the node appears under 🎨 DockerCPU API/🎨 Replicate. No models, no VRAM, nothing to cache.

Troubleshooting

• Search results you didn't ask for - the grounding features are the model's own opinions. If image_search pulls in context that distorts the output, switch it off and feed your own IMAGE_x references instead.
• Censorship, same as Pro - Google's IMAGE_SAFETY filtering applies across the Banana line. "Powerful, unimaginative, massive hidden biasing" is how one thread summed it up. If the model refuses a prompt, that's the platform, not your graph.
• Cost - pay-per-prediction. dry_run on the node skips the API call and returns a mock so you can check wiring before spending.
• Young pack - April 2026, zero stars, no community to lean on. The schema-driven approach works, but you're an early adopter.

The honest split: Pro if you need 4K output or maximum quality, Banana 2 if you want the search grounding and the panoramas, and honestly neither is the "daily driver local model" you already have on your GPU. This node is for the CPU-only box, or for the specific job where knowing what's on the web today matters more than staying off Google's cloud.