Extensions/ComfyUI-Gemini-3
ComfyUI Extension

ComfyUI-Gemini-3

A ComfyUI extension with 12 custom nodes.

By xuchenxu168·Created 9 months ago·Updated 9 months ago· 4
xuchenxu168/ComfyUI-Gemini-3
Nodes12
On cloudLocal install
CategoryGemini3/高级, Gemini3/优化
Stars4
Updated9 months ago

Nodes (12)

Gemini 3 聊天
Gemini3/高级
Gemini 3 创建缓存
Gemini3/优化
Gemini 3 函数调用
Gemini3/高级
Gemini 3 图像提示词编辑
Gemini3/实验性
Gemini 3 多模态分析
Gemini3/分析
Gemini 3 安全设置
Gemini3/安全
Gemini 3 流式生成
Gemini3/流式
Gemini 3 结构化输出
Gemini3/高级
Gemini 3 文本生成
Gemini3
Gemini 3 使用缓存
Gemini3/优化
Gemini 3 + 代码执行
Gemini3/工具
Gemini 3 + Google搜索
Gemini3/工具
Readme

ComfyUI-Gemini-3-2

Full-featured, high-performance Google Gemini 3 integration for ComfyUI. This plugin implements all Gemini 3 capabilities with optimized performance.

🔧 Latest Update (2025-11-19): Fixed Gemini API error 400 "Invalid thinkingLevel field". The thinkingLevel parameter is now correctly placed in the API request. See QUICK_FIX_GUIDE.md for details.

🌟 Features

Core Gemini 3 Capabilities

  • ✅ Dynamic Thinking Levels (Low/High) - Control reasoning depth vs latency
  • ✅ Media Resolution Control - Fine-grained control over image/video/PDF processing
  • ✅ Thought Signatures - Maintain reasoning context across API calls
  • ✅ Function Calling - Strict validation with parallel and sequential support
  • ✅ Structured Outputs - JSON schema validation with tools integration
  • ✅ Context Caching - Reduce costs for repeated context (2048+ tokens)
  • ✅ Batch API Support - Process multiple requests efficiently

Multimodal Support

  • 📝 Text generation with advanced reasoning
  • 🖼️ Image understanding and generation (up to 1120 tokens per image)
  • 🎥 Video analysis (70-280 tokens per frame)
  • 🎵 Audio understanding (speech, music)
  • 📄 PDF document processing (optimized at medium resolution)
  • 🔗 URL context and web search integration

Advanced Features

  • 🧠 Multi-step reasoning with thought signature preservation
  • 🔄 Streaming support with real-time updates
  • 🛡️ Safety settings with configurable thresholds
  • 📊 Token counting and usage tracking
  • 🌡️ Temperature control (optimized for Gemini 3)
  • 🔧 System instructions for behavior customization
  • 🔍 Google Search integration for real-time information
  • 💻 Code Execution for mathematical and data analysis tasks

📦 Installation

  1. Clone into your ComfyUI custom_nodes directory:
cd ComfyUI/custom_nodes
git clone https://github.com/yourusername/ComfyUI-Gemini-3-2.git
  1. Install dependencies:
cd ComfyUI-Gemini-3-2
pip install -r requirements.txt
  1. Set your API key:
    • Get your API key from Google AI Studio
    • Set environment variable: GEMINI_API_KEY=your_key_here
    • Or configure in the node settings

🎮 Nodes

基础节点 (Basic Nodes)

1. Gemini 3 文本生成 (Text Generation)

Generate text with advanced reasoning capabilities.

  • Inputs: prompt, images, audio, video, PDFs, thinking level, media resolution, system instruction
  • Outputs: text, thought signature, usage metadata
  • Features: Multimodal input, customizable thinking levels, media resolution control

高级节点 (Advanced Nodes)

2. Gemini 3 函数调用 (Function Calling)

Execute function calls with strict validation and thought signature preservation.

  • Inputs: prompt, function declarations, thought signatures, function responses
  • Outputs: function calls (JSON), updated thought signature, text response
  • Features: Parallel and sequential function calling, multi-step reasoning

3. Gemini 3 结构化输出 (Structured Output)

Generate JSON with schema validation and tool integration.

  • Inputs: prompt, JSON schema, enable Google Search/Code Execution/URL Context
  • Outputs: validated JSON output, usage metadata
  • Features: Strict JSON schema validation, integrated with built-in tools

4. Gemini 3 聊天 (Chat)

Multi-turn conversation with automatic context management.

  • Inputs: message, chat history, system instruction, max output tokens
  • Outputs: response, updated chat history (with thought signatures), usage metadata
  • Features: Automatic history management, thought signature preservation

流式节点 (Streaming Nodes)

5. Gemini 3 流式生成 (Streaming Generation)

Real-time streaming generation with progressive output.

  • Inputs: prompt, thinking level, system instruction, max output tokens
  • Outputs: full text, thought signature, stream chunks info
  • Features: Real-time response streaming, chunk-by-chunk processing

优化节点 (Optimization Nodes)

6. Gemini 3 创建缓存 (Create Context Cache)

Create cached contexts for cost optimization (10x cheaper for cached content).

  • Inputs: content to cache (min 2048 tokens), TTL minutes, cache name
  • Outputs: cache name/reference, cache info (with expiry), status
  • Features: Automatic TTL management, usage metadata tracking

7. Gemini 3 使用缓存 (Use Cached Content)

Generate content using previously cached context.

  • Inputs: prompt, cache name, thinking level
  • Outputs: response, usage metadata (showing cached token savings)
  • Features: Significant cost reduction, maintains full functionality

8. Gemini 3 批处理 (Batch Processing)

Process multiple requests efficiently with 50% discount.

  • Inputs: prompts JSON array, thinking level
  • Outputs: batch name/ID, status info
  • Features: 50% cost reduction, 24-hour processing window

工具节点 (Tool Nodes)

9. Gemini 3 + Google搜索 (With Google Search)

Generate content with real-time web search integration.

  • Inputs: prompt, thinking level, max output tokens
  • Outputs: response, grounding metadata (search sources), usage metadata
  • Features: Real-time information, automatic grounding, source attribution

10. Gemini 3 + 代码执行 (With Code Execution)

Generate content with automatic Python code execution.

  • Inputs: prompt, thinking level
  • Outputs: response, execution results (code + output), usage metadata
  • Features: Automatic code generation and execution, mathematical computations

安全节点 (Safety Nodes)

11. Gemini 3 安全设置 (Safety Settings)

Generate content with customizable safety thresholds.

  • Inputs: prompt, thinking level, harassment/hate speech/sexually explicit/dangerous content thresholds
  • Outputs: response, safety ratings
  • Features: Fine-grained safety control, detailed safety ratings, block reason reporting

🚀 Usage Examples

Basic Text Generation with High Thinking

Prompt: "Find the race condition in this multi-threaded C++ code..."
Thinking Level: High
→ Detailed analysis with deep reasoning

Image Analysis with High Resolution

Prompt: "Read all the text in this image"
Image: [your image]
Media Resolution: High (1120 tokens)
→ Accurate OCR and detailed analysis

Video Understanding

Prompt: "Describe what happens in this video"
Video: [your video]
Media Resolution: Low (70 tokens/frame) - for general description
Media Resolution: High (280 tokens/frame) - for text reading

Function Calling with Thought Signatures

1. Model calls check_flight → Returns signature A
2. Send flight result + signature A
3. Model calls book_taxi → Returns signature B
4. Send taxi result + signatures A & B
→ Model maintains reasoning chain

⚙️ Configuration

Thinking Levels

  • Low: Fast, minimal reasoning - for simple tasks
  • High (default): Maximum reasoning depth - for complex tasks

Media Resolution

  • Images: High (1120 tokens) recommended for most tasks
  • PDFs: Medium (560 tokens) optimal for documents
  • Video: Low/Medium (70 tokens) for general, High (280) for text

Temperature

  • Default: 1.0 - Optimized for Gemini 3
  • ⚠️ Changing temperature may cause looping or degraded performance

📊 Performance Tips

  1. Use Low Thinking for simple tasks to reduce latency
  2. Use Medium Resolution for PDFs (rarely benefits from High)
  3. Cache contexts 2048+ tokens for repeated use
  4. Batch requests when processing multiple items
  5. Keep temperature at 1.0 unless you have specific needs

🔗 Resources

📝 License

MIT License - See LICENSE file for details

🤝 Contributing

Contributions welcome! Please read CONTRIBUTING.md first.

💬 Support