🥭 Project Mango: Local Text-to-Image Inference Engine

Project Mango is a high-performance, CPU-optimized AI art generation suite. Unlike standard implementations that rely on high-level library wrappers, Mango features a raw PyTorch inference pipeline and a custom Server-Sent Events (SSE) streaming architecture to provide real-time feedback during the denoising process.

🛠️ Technical Architecture

• Custom Inference Loop: Built using discrete components (UNet, VAE, CLIP) to bypass standard black-box pipelines, allowing for manual tensor manipulation.
• Dynamic CFG Routing: Implements automated routing logic that dynamically adjusts the Classifier-Free Guidance (CFG) scale per request to vary output stylization (Dreamy, Standard, or Crunchy).
• Real-Time SSE Streaming: Uses Server-Sent Events to pipe mathematical progress from the Python backend directly to the frontend, enabling a live-updating progress bar.
• Decoupled Frontend: A modern, responsive interface built with a premium glassmorphism aesthetic, served via an optimized Flask routing system.

📂 Project Structure

mango/
├── frontend/
│   ├── static/          # CSS and JavaScript
│   │   ├── style.css
│   │   └── script.js
│   └── index.html       # Primary Application UI
├── output/              # Local directory for generated media
├── backend.py           # Flask server and SSE routing logic
├── mango.py             # Core PyTorch Diffusion Engine
└── README.md            # Project documentation

🚀 Getting Started

Prerequisites

• Python 3.8+
• RAM: 8GB minimum (16GB recommended)
• Storage: ~2GB for model weights

Installation

1. Clone the repository:

git clone https://github.com/omnimistic/mango.git
cd mango

2. Install dependencies:

pip install torch diffusers transformers flask pillow

3. Launch the Engine:

python backend.py

4. Access the UI: Open http://127.0.0.1:5000 in your web browser.

🌐 Network Access

By default, the server is locked to 127.0.0.1 (localhost) for security. To access the engine from other devices on your local network:

1. In backend.py, change host='127.0.0.1' to host='0.0.0.0'.
2. Navigate to your machine's local IP (eg: http://192.168.1.XX:5000) on your mobile device or tablet to access the interface.

⚖️ License & Acknowledgements

• The Code: Licensed under the MIT License.
• The Model: This project utilizes bk-sdm-tiny, a derivative of Stable Diffusion. Use of these weights is subject to the CreativeML Open RAIL-M license. Users are responsible for ensuring generated content complies with the ethical use-case restrictions outlined therein.