Nano Banana (Google Gemini Image)

"Nano Banana" started as a leaked codename on a HuggingFace leaderboard, got adopted by Google's own marketing, and is now the brand for Gemini's image models. This node is the original - Nano Banana (Gemini 2.5 Flash Image), Google's consumer-base image model, wired into ComfyUI as a built-in partner node. It's the model that made "type whatever, get a surprisingly good image" feel effortless, and it's also the cheap rung: roughly 6 credits an image, which is less than a third of what the Pro tier costs.

Its real strength is editing with context. You don't get a mask or a denoise slider - you get conversation. Attach one or more images, type "make it snow but keep the people exactly as they are," and the model reasons over the whole scene. That's the Gemini way, and it's why people reach for this node when local diffusion editing (mask, inpaint, composite) feels like surgery.

How it works and the inputs that matter

The node sends your prompt, optional reference images/files, aspect ratio, and a system prompt to Gemini's generate-content endpoint and gets back an image - no local model, no VRAM. Settings that matter:

• prompt - the instruction or description. This is a reasoning model, so conversational phrasing works: "a fox in a spacesuit, editorial photo, dramatic rim light" or "change the sign to say COFFEE."
• images - optional reference images (use the Batch Images node for several). This is the editing path; with none, it's text-to-image.
• aspect_ratio - defaults to auto: match your input image, or 1:1 squares when there's no image. Otherwise pick from 1:1 to 21:9.
• response_modalities - IMAGE or IMAGE+TEXT. The latter returns the image and a text response (a second STRING output), which is great when you want the model to also caption its own work.
• system_prompt - an advanced override that comes pre-filled with Google's "you are an expert image-generation engine" instructions. You can replace it; you usually shouldn't need to.
• seed - best-effort reproducibility, "not guaranteed," per its own tooltip. Treat it as "vary the output" rather than "lock the output."

Cost, setup, and what to watch

Ships with ComfyUI core; needs a Comfy account with credits and a current ComfyUI; login from localhost (or via the Comfy API key on a LAN). Cost is about $0.04 per 1K-image - cheap enough for real iteration, which is exactly how people use it: generate a look here, refine locally, upscale locally, and only pay when the creative decisions are basically made.

The two honest caveats: it's a hosted service, so your images go to Google and the content filter is Google's (this model got a reputation for heavy and at times unpredictable safety filtering - the "unimaginative, massive hidden biasing" complaint is real). And if you need 4K output or fourteen reference images, that's the Nano Banana Pro node (GeminiImage2Node) - this one tops out at 1K and single-digit references. Pick the banana to fit the job.