ComfyUI-Nano-Banana

Custom node pack for Nano Banana (Gemini image generation and editing) in ComfyUI.

Nodes

| Node | Description | |------|-------------| | Google API Key | Provides API key via env var, config file, or manual input | | Nano Banana Generate | Text-to-image and image-to-image via Gemini models (:generateContent endpoint) | | Nano Banana Chat Edit | Multi-turn conversational image editing and transformation |

Installation

1. Clone into ComfyUI/custom_nodes/:

   cd ComfyUI/custom_nodes
   git clone https://github.com/fwlemos/NBP-Comfyui.git
   

2. Install dependencies:

   pip install -r NBP-Comfyui/requirements.txt
   

3. Restart ComfyUI.

API Key Setup

The extension resolves your Google API key using the following priority order:

1. Environment Variables: GOOGLE_API_KEY or GEMINI_API_KEY.
2. Config File: A google_api_key.txt file inside the extension directory.
3. Node Input: The api_key input on generation/chat nodes (or the dedicated Google API Key node).

[!NOTE] This priority ensures you can have a global key set but still override it for specific workflows.

Nano Banana Generate

Generates images using Google's Gemini vision-language models via the :generateContent endpoint.

Models

• gemini-3-pro-image-preview (Nano Banana Pro) — Highest quality, best text rendering.
• gemini-3.1-flash-image-preview (Nano Banana 2 / Flash) — Faster, cheaper, supports reasoning/thinking & 512px.

Parameters

Generation & Sampling

• prompt: Text description of the image. Supports multiple languages.
• model: Gemini model selection (Pro or Flash).
• aspect_ratio: Output proportions (Up to 14 ratios supported on Flash).
• image_size: Resolution (512px, 1K, 2K, 4K). Note: 512px is Flash-only.
• temperature: Controls randomness (0.0–2.0). 1.0 is recommended for Gemini 3.
• top_p / top_k: Nucleus and Top-K sampling parameters for variety control.
• seed: Sampling seed (0 = random). Influences consistency but doesn't guarantee identical results.

Logic & Batching

• reference_images (Optional): Connect up to 14 images for image-to-image or identity locking.
• batch_count: Number of concurrent API calls (1–8).
• candidate_count: Variations per API call (1–4). Total images = batch_count × candidate_count.
• max_output_tokens: Total token budget (text + thinking + image). Recommended: 2048+.
• thinking_level: Reasoning depth (Flash only). Options: minimal, low, medium, high.
• response_modality: IMAGE (extracts only imagery) or TEXT_AND_IMAGE (includes model descriptions).
• enable_search_grounding: Uses Google Search to inform generation with real-world knowledge.

Safety & Filters

• safety_...: Configurable filters for Hate Speech, Harassment, Sexually Explicit, and Dangerous Content.
• stop_sequences: (Optional) Strings that trigger immediate generation termination.

Nano Banana Chat Edit

A conversational interface for iterative image editing.

Parameters

• instruction: The edit command (e.g., "Change the sky to sunset").
• input_image: The base image to be transformed.
• chat_history: Chain multiple chat nodes to maintain context across turns.
• system_instruction: High-level persona or style guidance for the model.
• Includes all sampling and safety parameters from the Generate node.

License

MIT