FAL Nano Banana

Nano Banana is Google's rebrand of Gemini-native image generation - the name started as an anonymous codename on a public model leaderboard, Google's own people started dropping banana emojis about it on social media, and the joke just became the official product name. This node calls it through FAL's hosted API rather than Google's directly, giving you an image-editing model in your graph that isn't a diffusion checkpoint at all - it's a large multimodal model that reasons about the image and prompt together, which is a meaningfully different tool than a Stable-Diffusion-family editor.

What it actually does: you supply a prompt describing the edit and up to four reference images (or a list of image URLs instead), and it returns an edited result. model is a required enum with two choices: fal-ai/nano-banana/edit (the consumer-tier "Nano Banana," roughly Gemini 2.5 Flash Image under the hood) and fal-ai/nano-banana-pro/edit (the flagship "Pro" tier, Gemini 3 Pro Image class - natively capable of 4K output, more reference images, and a "thinking mode" that reasons about the prompt before generating, at a meaningfully higher per-image cost than the base tier).

Inputs that matter: prompt is required. image through image_4 are optional individual image inputs - up to four references - while image_urls is an alternative multiline field if your references are already hosted rather than loaded locally. num_images (1–10, default 1) generates multiple variations in one call. aspect_ratio defaults to auto with eleven presets available from 21:9 down to 9:16. resolution (1K/2K/4K, default 1K) is where the Pro tier's 4K capability actually gets used - the base nano-banana/edit model may not honor higher resolutions the same way the Pro variant does, so match your model choice to the resolution you actually need. output_format (jpeg/png/webp, default png) and sync_mode round out the request-shaping options.

Outputs: image, width, height, and description - the last one is worth noticing, since it means this model can return text describing what it did or what it sees, on top of the edited image itself, which is a genuinely different output shape than a plain diffusion editing node gives you.

Install: search "ComfyUI-utils-nodes" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/zhangp365/ComfyUI-utils-nodes

Restart afterward. You need a FAL API key from fal.ai's account settings, via the api_key input or a fal_config.yml file.

The thing to actually know before you use this:

• Google's image models have a real reputation for aggressive, sometimes opaque content filtering - the community's read on the wider Nano Banana lineup has been "powerful, unimaginative, massive hidden biasing." If a generation comes back refused, blank, or oddly conservative compared to what a diffusion model would have done with the same prompt, that's Google's safety layer, not a bug in this node - and it can trip on things that aren't obviously restricted at first glance. Google also tightened restrictions specifically around recognizable public figures and celebrity likeness in a January 2026 policy update, so expect edits involving real or real-seeming people to be the most filter-prone case.
• All Nano Banana output carries an invisible SynthID watermark baked in by Google at generation time. That's not something this node or FAL can strip or opt out of - it's a property of the underlying model, worth knowing if provenance-clean output matters for your use case.
• Pro-tier calls at 4K cost meaningfully more per image than the base tier at 1K - sanity-check your prompt and framing on the cheaper nano-banana/edit model at 1K before switching to nano-banana-pro/edit at higher resolutions for a final pass.
• timeout defaults to 300 seconds; multi-image, higher-resolution Pro requests are the ones most likely to need it raised toward the 3000-second ceiling.