ComfyUI_Gemini3ProImagePreview

🍌 NanoBanana2 – Gemini Image Node for ComfyUI

NanoBanana2 is a custom ComfyUI node for Google Gemini image generation models.

Supported models:

• gemini-3.1-flash-image-preview (Nano Banana 2)
• gemini-3-pro-image-preview
• gemini-2.5-flash-image

Features:

• Text → Image
• Image → Image (edit / refine)
• Multi-turn refinement support
• Resolution control (0.5K / 1K / 2K / 4K)
• Extended aspect ratios (Nano Banana 2)
• Optional Google Search grounding
• Deterministic seed control
• Local resolution enforcement (pixel-budget based resizing)

---

Installation

1. Install into ComfyUI

Clone or copy this repository into:

ComfyUI/custom_nodes/NanoBanana2

Restart ComfyUI after installation.

---

2. Install Requirements

Activate your ComfyUI environment and run:

pip install -r requirements.txt

Minimum dependencies:

aiohttp pillow torch numpy

---

API Setup

NanoBanana2 requires a Google Gemini API key.

1. Create API Key

Visit: https://ai.google.dev/

Create a Gemini API key.

2. in .env file change GEMINI_API_KEY=YOUR_API_KEY to your actual API-key.

OR

3. Set Environment Variable

Windows (PowerShell):

setx GEMINI_API_KEY "YOUR_API_KEY"

Restart terminal and ComfyUI after setting it.

macOS / Linux:

export GEMINI_API_KEY="YOUR_API_KEY"

To verify:

echo $GEMINI_API_KEY

---

Supported Models

Model | 0.5K | Extended Aspects | Thinking Config | Image Search

gemini-3.1-flash-image-preview | Yes | Yes | Yes | Yes gemini-3-pro-image-preview | No | Limited | No | Web only gemini-2.5-flash-image | No | Limited | No | Web only

---

Node Settings

Model Select Gemini image model. Nano Banana 2 exposes the most features.

Aspect Ratio Includes auto, standard (1:1, 16:9, 9:16, 3:2, etc.), and extended ratios on Nano Banana 2. Unsupported ratios are sanitized automatically.

Resolution Resolution is interpreted as pixel area budget:

0.5K → ~512 × 512 area (~262K pixels) 1K → ~1024 × 1024 area 2K → ~2048 × 2048 area 4K → ~4096 × 4096 area

For non-square ratios: width × height ≈ base_size²

Example: 0.5K + 16:9 → approximately 688 × 384

The node enforces resolution locally to guarantee correct output size even if Gemini returns original input dimensions during edits.

Seed -1 → random Any positive integer → deterministic output

Temperature 0.2 – 0.6 → Stable 0.7 – 1.2 → More creative variation

Google Search Optional grounding for improved realism or current references. Nano Banana 2 additionally supports image search grounding.

Thinking Level (Nano Banana 2 only) minimal High

Higher thinking levels increase reasoning effort before generation.

Image Input (Optional) If an image is connected:

• Performs edit/refinement
• Resolution is enforced locally
• Multi-turn refinement is preserved

---

Multi-Turn Refinement

The node preserves chat metadata to support:

• Iterative refinement
• Context-aware editing
• Image evolution workflows

---

Notes

• Unsupported settings are automatically sanitized.
• Extreme aspect ratios may reduce composition quality.
• Output resolution is always enforced locally to match selected pixel budget.

---

License

Add your preferred license here.