Z-Image-Turbo

✨ Update: MCP Server support! Setup Guide

A professional web interface for the Tongyi-MAI Z-Image-Turbo model — lightning-fast text-to-image generation with 6B parameters.

🚀 Quick Start

Prerequisites

• Python 3.8+
• Node.js 16+
• 8GB+ VRAM recommended (or use CPU offload)

Installation

1. Clone the repository

   git clone https://github.com/Aaryan-Kapoor/z-image-turbo.git
   cd z-image-turbo
   

2. Backend Setup

   python -m venv venv
   
   # Windows
   .\venv\Scripts\activate
   # Linux/Mac
   source venv/bin/activate
   
   pip install -r backend/requirements.txt
   

3. Frontend Setup

   cd frontend
   npm install
   

Running the Application

Terminal 1 - Start Backend:

.\venv\Scripts\activate  # or source venv/bin/activate on Linux/Mac
cd backend
python main.py

Terminal 2 - Start Frontend:

cd frontend
npm run dev

Open http://localhost:5173 in your browser and start generating!

🔌 MCP Server (Model Context Protocol)

Please refer to MCP_README.md for a full guide on implementation with LM Studio and Claude Desktop.

Z-Image-Turbo now includes a powerful MCP server that exposes image generation capabilities through the standardized Model Context Protocol. This allows AI assistants (like Claude), automation tools, and other MCP-compatible clients to generate images programmatically.

Why Use the MCP Server?

• AI Integration: Let Claude or other AI assistants generate images directly during conversations
• Automation: Build automated workflows that include image generation
• Remote Access: Generate images from web clients or remote services (HTTP mode)
• Standardized API: Use the same protocol across different AI tools and platforms

Quick Start with MCP

1. Install MCP dependencies:

cd backend
pip install -r requirements.txt

2. Run the MCP server:

For local integration (Claude Desktop, MCP Inspector):

cd backend
./run_mcp.sh --stdio

For HTTP/web clients and remote access:

cd backend
./run_mcp.sh --http --port 8001
# Server available at http://localhost:8001/mcp

3. Configuration: Edit backend/mcp_config.json to set default transport mode and port:

{
  "transport": "stdio",
  "host": "0.0.0.0",
  "port": 8001
}

Available MCP Tools

⚙️ Production Configuration

Edit backend/mcp_config.json to customize server behavior:

{
  "transport": "stdio",
  "eager_load": false,
  "model_ttl_minutes": 0,
  "max_concurrent_requests": 1,
  "log_level": "INFO"
}

Usage Example

Once connected to Claude Desktop or another MCP client:

You: "Generate an image of a serene mountain landscape at sunset"

Claude: [Uses generate_image tool]
{
  "prompt": "A serene mountain landscape at sunset with vibrant orange and purple skies",
  "width": 1024,
  "height": 768,
  "num_inference_steps": 8
}

[Returns rendered image]

Transport Modes Comparison

LM Studio Integration

Add to your LM Studio MCP config file:

{
  "mcpServers": {
    "z-image-turbo": {
      "command": "C:\\path\\to\\z-image-turbo\\venv\\Scripts\\python.exe",
      "args": [
        "C:\\path\\to\\z-image-turbo\\backend\\mcp_server.py",
        "--transport",
        "stdio"
      ],
      "env": {
        "PYTHONUNBUFFERED": "1"
      },
      "timeout": 300000
    }
  }
}

Important:

• Replace C:\\path\\to\\z-image-turbo with your actual installation path
• Use double backslashes \\ for Windows paths in JSON
• Point to the venv Python executable (not system Python!)
• Model loads on first request by default (lazy loading)
• Add --eager-load to args if you want model to load at startup
• timeout: 300000 (5 minutes in ms) for model loading + generation

Claude Desktop Integration

Add to your Claude Desktop config file (~/Library/Application Support/Claude/claude_desktop_config.json on macOS or %APPDATA%\Claude\claude_desktop_config.json on Windows):

{
  "mcpServers": {
    "z-image-turbo": {
      "command": "C:\\path\\to\\z-image-turbo\\venv\\Scripts\\python.exe",
      "args": [
        "C:\\path\\to\\z-image-turbo\\backend\\mcp_server.py",
        "--transport",
        "stdio"
      ],
      "env": {
        "PYTHONUNBUFFERED": "1"
      },
      "timeout": 300000
    }
  }
}

⚠️ Critical:

• Use the venv Python path, not "python" (system Python won't have dependencies!)
• Use double backslashes \\ for Windows paths in JSON
• Model loads on first request by default (saves memory)
• Add --eager-load to args if you want model to load at startup (avoids first-request timeout)
• timeout: 300000 (5 minutes) ensures enough time for model loading

After restarting Claude Desktop, you can ask Claude to generate images and it will use the MCP server automatically!

Testing Your MCP Server

Test with the official MCP Inspector:

npx @modelcontextprotocol/inspector python backend/mcp_server.py --transport stdio

This opens a web interface where you can test all available tools and inspect requests/responses.

📖 Full MCP Documentation: See MCP_README.md for detailed setup, troubleshooting, and complete deployment guide.

✨ Features

Application

• Premium Dark UI — Glassmorphism design with intuitive controls
• Smart Presets — Quick aspect ratios (1:1, 3:4, 16:9) and resolutions (480p-1080p)
• Fine Control — Sliders for dimensions, inference steps, guidance scale, and seed
• Real-time Progress — Live generation tracking
• Flexible Deployment — Custom model cache directory, CPU offload option

MCP Server Integration

• 🔌 Dual Transport Modes — Support for both stdio (local) and HTTP/SSE (remote) connections
• 🤖 AI Assistant Compatible — Seamless integration with Claude Desktop and other MCP clients
• 🛠️ Rich Tool Set — Image generation, model info, configuration management, and example prompts
• ⚙️ Configurable — Customizable host, port, and transport settings via mcp_config.json
• 🔒 Production Ready — Stateless HTTP mode for scalable deployments

Model (Z-Image-Turbo)

• ⚡ Lightning Fast — Optimized for 8-step generation, achieving sub-second latency on enterprise GPUs.
• 🏗️ S3-DiT Architecture — Built on Scalable Single-Stream Diffusion Transformer technology.
• 🧠 Advanced Encoders — Uses Qwen 4B for powerful language understanding and Flux VAE for image decoding.
• 🎓 DMDR Training — Trained using Fusing DMD with Reinforcement Learning for superior semantic alignment.
• 🌐 Bilingual Mastery — Exceptional rendering of text in both English and Chinese.
• 🎨 Versatile & Uncensored — From photorealism to anime, handling complex concepts without censorship.
• 📐 High Fidelity — Native support for resolutions up to 2MP (e.g., 1024x1536, 1440x1440).
• 💾 Efficient — 6B parameters, comfortably fitting in 16GB VRAM (consumer-friendly).

🔬 Technical Architecture

Z-Image-Turbo represents a significant leap in efficient generative AI:

• Base Architecture: S3-DiT (Scalable Single-Stream DiT)
• Text Encoder: Qwen 4B (Large Language Model based conditioning)
• VAE: Flux Autoencoder
• Training Method: Distilled from Z-Image using DMDR (DMD + RL)
• Inference: 8 NFEs (Number of Function Evaluations) default
• Precision: Optimized for bfloat16 / fp8

🛠️ Tech Stack

• Backend: FastAPI, PyTorch, Diffusers, Transformers
• Frontend: React, Vite, Lucide React
• MCP Server: FastMCP, Starlette (supports stdio and HTTP/SSE transports)
• Model: Tongyi-MAI/Z-Image-Turbo (6B parameters)

⚙️ Configuration

Access settings via the gear icon in the sidebar:

• Model Cache Directory — Specify where to download/store the model
• CPU Offload — Enable for GPUs with limited VRAM

📝 License

This project is open-source under the Apache 2.0 License.

🙏 Credits

• Model: Tongyi-MAI/Z-Image-Turbo by Alibaba Group
• UI Framework: React + Vite
• Backend: FastAPI + Diffusers