NanoBanana Full API

Nano Banana - Google's Gemini image model - has no open weights. You can't run it locally, so the only way to get it inside a ComfyUI workflow is through an API node that calls it for you. That's exactly what this node is: an HTTP client dressed as a generator. You type a prompt, it talks to Google's generateContent endpoint, and the returned image lands on a normal IMAGE socket that feeds the same preview, upscale, and save nodes as your local output. From the canvas it looks like any sampler. Underneath it's a network call with your API key.

The class name (NanoBanana_Gemini_2_5_Flash_V2) is a historical artifact - the node keeps the old workflow ID for compatibility, but it's been updated to Google's current endpoints and models. You get two model choices in model_version:

• gemini-3.1-flash-image (default) - the general-purpose Nano Banana successor.
• gemini-3-pro-image - the premium tier for complex instructions and professional asset work. Note: thinking_level is a Flash-only control; Pro manages its own thinking.

Keys and credentials

The key resolves in order: the api_key widget, then the GEMINI_API_KEY env var, then GOOGLE_API_KEY. It's sent in the x-goog-api-key header, never the URL or body. Two practical notes: keys typed into the widget get serialized into the workflow JSON, so use the environment variable for anything you'll share; and the node surfaces empty-credential, provider, timeout, and safety errors instead of dying silently.

The inputs that actually matter

• prompt - the instruction. system_prompt - only prepended if you change it from the default.
• aspect_ratio and resolution (1K/2K/4K) - the output shape. 4K default is greedy; 1K is plenty for drafts.
• seed, temperature, top_p - generation controls you mostly leave alone until you're iterating.
• candidate_count (1–10) - how many candidates to ask for per call.
• safety_harassment / safety_hate_speech / safety_sexual / safety_dangerous - per-category blocking policy (default BLOCK_DEFAULT). Google enforces its own content rules at the source regardless; this only tunes how aggressively the built-in filter blocks.
• edit_mode_enabled - set to yes to treat the optional mask as a mask for image 1, for image editing.
• Optional image_1…image_5 (IMAGE) - reference images; up to five for editing or composition.
• show_thoughts and thinking_level - request/report the model's reasoning.

Outputs: image_batch (IMAGE) - the returned candidates - and text_output (STRING) - provider text or response details, including thoughts if you asked for them.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/haroonaslam/WorkflowX-Configurator

or search WorkflowX Configurator in ComfyUI Manager, restart, refresh. No pip deps and no model download - the "model" lives on Google's servers.

The honest tradeoffs

• Per-call cost, and it adds up. Every generation is a metered API call. Draft at 1K, not 4K.
• Your inputs leave the machine. Prompt and reference images go to Google. If that's a dealbreaker, this node isn't for you - it's the mechanism, not a bug.
• Filtering follows the model. You cannot abliterate a closed model's refusals; whatever Google refuses, this node refuses.
• It's a third-party pack with your key. That's the category that once shipped credential-stealing malware (ComfyUI_LLMVISION), so install it via Manager from the registry and keep the key in an env var - not pasted into every workflow you share.

The cleanest way to get Nano Banana-class generation into a local workflow, with the API tax fully in view.