RunPod Nano Banana Edit (Google)

"Nano Banana" is the community's name for Google's Gemini image models, and the whole demand-side story is one sentence long: people want the closed Google image editor inside their workflow, next to their local upscaler and their masking. This node is exactly that - nano-banana-edit as a RunPod public endpoint, taking two image inputs (input_image_1, input_image_2) and a text instruction, and returning the combined result as an IMAGE tensor. It's the pack's take on the most-hyped closed image model of the last couple of years.

The hype is partly deserved - combining two images into one coherent scene ("merge this person and this background", "blend these two subjects") is something Google's model does frighteningly well. But it's also the most resold, proxied, and marketed closed model in the space, which is worth keeping in mind: you're paying a third-party pipeline (this node → RunPod → the model) for a service that a reseller layer also sells at a near-identical price. The value here is purely that it's in ComfyUI, on your canvas, feeding your existing workflow.

How it works

Async POST to nano-banana-edit, poll until done, base64 image back out. Both wired input images become data URLs collected into the images array in the payload (merged with whatever URLs you put in the images JSON field). Note this node has no size and no seed inputs - the endpoint doesn't take them. What you get instead is prompt + images + a safety toggle. Keep that in mind when you're tweaking for reproducibility.

The inputs that matter

• input_image_1 / input_image_2 - the two sources to combine. Wire from Load Image nodes.
• prompt - the instruction: "Combine these images into a single realistic scene" is the pack's default, and realistic is the operative word - this model shines at making two inputs feel like one photograph.
• images - JSON array of URLs, for hosted references instead of wires.
• enable_safety_checker - on by default; Google's moderation applies at the source and there's no bypass.

What comes out

One image output (IMAGE). Save, preview, upscale - normal tensor behavior.

Installing it

One of 27 nodes in tcarwash/ComfyUI_RunpodNodes. ComfyUI Manager: search ComfyUI_RunpodNodes, or

cd ComfyUI/custom_nodes
git clone https://github.com/tcarwash/ComfyUI_RunpodNodes

then restart. No requirements.txt - the pack runs on what ComfyUI already ships (requests, Pillow, torch). Grab a RunPod key at runpod.io.

Gotchas

• No seed, no size - if you're coming from local generation this feels like a missing limb. It's the endpoint's design; you just re-run until you like it.
• Closed + metered + filtered - the full closed-model stack: no weights, per-image billing, moderation you can't remove.
• Queue waits - 30–120 seconds of polling; not hung.
• "Job failed" - key or credits, the pack's standard first checks.

Nano Banana is the most overhyped closed image model of its era, but hype aside, the two-image combination genuinely works. This node is the cleanest way to get that specific trick inside ComfyUI without a Google Cloud billing account.