Gemini Nano Banana Image

The name is the whole pitch. "Nano Banana" is Google's branding for its Gemini-native image generation - the same model behind ImageFX - and this node drops it straight into your ComfyUI graph. No checkpoint to download, no VRAM to budget, no LoRA training. You type a prompt, pay Google per image, and the result lands in a normal IMAGE tensor like any other sampler output.

The trade is obvious, so let's name it up front: this is an API wrapper, not a model. Everything happens on Google's servers. That buys genuinely strong text rendering and reference-image editing, and it costs you the thing most of ComfyUI's ecosystem is about - local, uncensored, free iteration. If you're after anime LoRAs or NSFW work, stop reading; the filters are baked in on Google's side and no node setting undoes them. If you need clean infographics, logos, product shots, or strong image-to-image edits, this is a lot of capability for a one-line dependency.

How it works

Under the hood the pack is refreshingly honest: it uses Google's official google-genai Python SDK and calls client.models.generate_content() with a GenerateContentConfig. That's it. The node's whole job is translating ComfyUI tensors into what the API wants and back - reference images become base64 PNGs, and the response's inline image data gets decoded back into a batched IMAGE tensor and a STRING of the model's text reply.

There are no hidden model files. requirements.txt is a single line: google-genai>=1.0.0. The models themselves are the three presets: gemini-2.5-flash-image (the original Nano Banana), gemini-3.1-flash-image-preview (Nano Banana 2), and gemini-3-pro-image-preview (Nano Banana Pro, the 4K flagship). Google rotates these names, which is why there's a custom_model field - type any newer model string there and it overrides the dropdown.

The inputs that actually matter

There are a lot of knobs, but a beginner really touches four:

• prompt - your text-to-image or edit instruction. Multi-line, do everything here.
• api_key / api_key_env - resolution order is api_key, then api_key_env (defaults to GEMINI_API_KEY), then GOOGLE_API_KEY. Paste a key in the node or set an env var and leave it alone; the env route is better if you share workflows.
• model - pick your tier. Pro is the one people rave about; Flash is cheap for iterating.
• seed - -1 (the default) means "unspecified." Worth knowing Google's seeds are less deterministic than a local sampler's; don't expect pixel-perfect reproduction across calls.

The useful middle shelf: aspect_ratio (1:1 up to 21:9), image_size (1K/2K/4K), candidate_count (request 1–4 and pick), response_mode (TEXT_AND_IMAGE vs IMAGE_ONLY), and person_generation, which you may need to set to ALLOW_ADULT just to get a normal person in frame. enable_google_search lets Gemini browse before generating - flashy, rarely worth it. proxy_mode matters behind a proxy: the SDK honors environment proxies unless you set it to direct.

For image-to-image and editing, wire up to five reference images (image_1 through image_5) into the optional inputs and describe the change. One gotcha from the source: gemini-2.5-flash-image is capped at 3 reference images on this node; the other models take the full 5. And if Google returns mixed-size images, the node keeps only the first size group and appends a note to the text output - so keep batch sizes consistent.

The outputs are image (an IMAGE tensor - feed it to SaveImage, an upscaler, whatever) and text (the model's text reply, which can carry the mixed-size note). If the API returns no image at all, the node raises a hard error with the raw response text - Google's way of telling you a safety filter tripped or the model name is wrong.

Installing it

ComfyUI Manager is the easy route - search for comfyui-nano-banana-custom-node and install. Or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/Leee62/comfyui-nano-banana-custom-node

Then install the SDK in the same Python environment ComfyUI runs in:

pip install -q -U google-genai

Restart ComfyUI and the node appears under Google AI/Gemini as "Gemini Nano Banana Image." No model download step - the only setup is the key.

Where people get burned

The #1 complaint you'll find in the wild isn't a bug in this node, it's the bill. Nano Banana's API is not free, and the free tier evaporates fast while you experiment. A graph that loops or a stray candidate_count of 4 multiplies your cost without multiplying your luck. Watch your billing page when you're tuning prompts.

Second: censorship. Google's image safety filtering is aggressive, and "powerful, unimaginative, massive hidden biasing" is the community's honest summary. If a generation silently returns nothing or a refusal, that's Google's filter, not your install - check the text output for the reason.

Third: don't paste your API key into shared workflow JSON. The api_key_env route exists precisely so a public workflow doesn't leak your billing account. Use it.