FLUX.2 Klein 4B — Local Image Generation

A fully local AI image generation app powered by FLUX.2-klein-4B, running on your Mac (Apple Silicon / MPS) or CPU. Includes a Flask API backend, a React frontend, and an MCP server for Claude Code integration.

Features

• Text-to-image generation
• Image-to-image and multi-reference modes (up to 4 reference images)
• Configurable resolution, steps, guidance scale, seed, and batch size (up to 4 images)
• GGUF-quantized model (Q4_K_M by default, ~2.6 GB) for low memory usage
• All generated images saved locally to outputs/ with metadata
• MCP server exposes generation tools directly to Claude Code

Requirements

• Python 3.10+
• Node.js 18+
• ~3–8 GB free RAM (depending on quantization level)
• Apple Silicon recommended (MPS acceleration); CPU fallback supported

Setup

bash setup.sh

This will:

1. Create a Python virtual environment and install dependencies
2. Install frontend dependencies and build the React app

The GGUF model (~2.6 GB for Q4_K_M) is downloaded automatically from HuggingFace on the first generation request.

Running

Production (serves the built React frontend):

source venv/bin/activate
python app.py
# Open http://localhost:5001

Development (hot reload):

# Terminal 1
source venv/bin/activate && python app.py

# Terminal 2
cd frontend && npm run dev
# Open http://localhost:3000

Quantization

Control memory usage via the FLUX_QUANT environment variable:

FLUX_QUANT=Q8_0 python app.py

API

Generate request body:

{
  "prompt": "a cat on the moon",
  "mode": "txt2img",
  "width": 768,
  "height": 768,
  "steps": 4,
  "guidance_scale": 4.0,
  "seed": null,
  "num_images": 1,
  "images": []
}

MCP Integration

The MCP server (mcp_server.py) exposes image generation tools to Claude Code. Configure it in .mcp.json to use the tools directly from your Claude session.

License

MIT