☁️BizyAir NanoBanana

Nano Banana is Google's branding for Gemini-native image generation - the name started as an anonymous codename on a leaderboard, Google's own people started dropping banana emoji about it on social media, and the meme name just became the official one. This node routes to gemini-2.5-flash-image, the base tier (not Nano Banana Pro's 4K/14-reference flagship), through BizyAir's cloud proxy - so you're not creating a Google Cloud project or fiddling with Gemini API keys, just spending BizyAir credits.

Why an image node has temperature and top_p

Because Gemini's image generation isn't a diffusion model with a text encoder bolted on - it's a multimodal LLM that reasons about the prompt as language before it renders anything. That's why this node exposes temperature and top_p (both 0–1, default 1) and max_tokens (up to 8192) - those are LLM sampling knobs, not diffusion parameters, and they're here because the underlying model genuinely works that way. If those settings look out of place on an image node, that's not a mistake in the pack, that's what Nano Banana actually is under the hood.

What it does

prompt plus operation - a four-way switch between generate (from nothing), edit (modify a reference), style_transfer, and object_insertion. The five optional image through image5 slots are your references - how many you actually need depends on which operation you picked; generate can ignore them entirely, while edit/style_transfer/object_insertion want at least one.

quality (standard/high, default high) and aspect_ratio (1:1, 16:9, 9:16, 4:3, 3:4, or auto) are the straightforward output controls. character_consistency (default on) is worth knowing about explicitly - it tries to keep a subject's identity stable across an edit chain, which matters if you're iterating on the same character across multiple edit calls and don't want the face drifting each pass.

seed runs 0 up to a fixed max for reproducibility.

Outputs

image (the result), a string (the model's accompanying text response - useful if you want to see what Gemini "said" about the generation, since it's reasoning in language before producing pixels), and the bizyair_model_name passthrough label.

Installing it

ComfyUI Manager: search BizyAir, install, restart. Or:

cd ComfyUI/custom_nodes && git clone https://github.com/siliconflow/BizyAir.git

Restart afterward. Nothing downloads locally - this is a thin client to Google's hosted model via BizyAir's account layer. You need a BizyAir account with an API key set before this node produces anything; the pack prompts you through it on first use.

Where people get burned

Expecting Nano Banana Pro's specs from the base model. This node targets the Flash tier, not the Pro flagship - no native 4K, no "thinking mode," no 14-reference-image ceiling. If you need those, you're looking for a different node in a different pack.

Google's filtering is aggressive and largely opaque. Community reports on Nano Banana describe it as technically strong but heavily biased toward "safe" outputs, and January 2026 policy changes tightened restrictions further, especially around recognizable people and IP. If a request silently fails or comes back nothing like what you asked for, filtering - not the node - is the likely cause.

Leaving character_consistency on when you don't want it. If you're deliberately trying to restyle a subject into something different across edits, this default can actively fight you.