Fast Style Transfer - Android App

A native Android application implementing Fast Neural Style Transfer for artistic image transformation on mobile devices. Built with custom TensorFlow optimizations and efficient on-device inference.

📱 Demo Download

Download APK - Working demo for Android devices

Note: Enable "Install from unknown sources" in your Android settings

✨ Features

• Real-time Style Transfer: Transform photos with artistic styles inspired by famous painters
• Multiple Art Styles: Van Gogh, Picasso, and other artistic styles available
• 100% On-Device Processing: All inference runs locally - no network required, no data uploaded
• Dynamic Input Support: Handles variable image sizes without predefined tensor dimensions
• Custom Model Optimization: Hand-tuned for mobile hardware constraints
• Intuitive UI: Clean, user-friendly interface with gallery integration
• Performance Optimized: ~4-5 second inference time on modern devices

🖼️ App Screenshots & Flow

Main Interface

Gallery Open

Photo Selection

Style Selection

Processing

Results

Inference time: ~4.3 seconds on test device

🛠️ Technical Implementation

Architecture Overview

• Framework: Native Android with TensorFlow Lite
• Model: Custom Fast Neural Style Transfer implementation
• Optimization: TensorFlow C_API integration for performance
• UI: Modern Android Architecture Components

Key Technical Features

• Custom Architecture: Redesigned Fast Neural Style Transfer network optimized for mobile deployment
• Ultra-Lightweight Model: Reduced to only 210,000 parameters through architectural innovations
• Dynamic Input Handling: Supports variable image dimensions without predefined tensor sizes
• Custom TensorFlow Inference: Hand-written inference pipeline bypassing TensorFlow Lite API limitations
• 100% Offline Processing: Complete on-device inference - no network connection required
• Individual Style Training: Each artistic style trained from scratch for optimal quality
• Android NDK Integration: Native C/C++ components for performance-critical operations
• Privacy-First Design: Zero data upload - all processing stays on your device

Performance Optimizations

• Model size optimization for mobile deployment
• Custom inference pipeline reducing overhead
• Memory-efficient image processing
• Hardware-specific optimizations

🔧 Technical Stack

• Languages: Java, C++, Python (training)
• Frameworks: Android SDK, TensorFlow, TensorFlow Lite
• Native Development: Android NDK, JNI
• Model Training: TensorFlow (Python)
• Build Tools: Gradle, CMake

📱 System Requirements

• Android Version: 6.0 (API level 23) or higher
• RAM: Minimum 3GB recommended
• Storage: 50MB for app + processing space
• Permissions: Camera, Storage access

🚀 Installation & Usage

From APK

1. Download the APK from the link above
2. Enable "Install from unknown sources" in Android settings
3. Install the downloaded APK
4. Grant necessary permissions (camera, storage)

Using the App

1. Gallery: Select a photo from your device or take a new one
2. Styles: Choose from available artistic styles
3. Transfer: Apply the selected style to your image
4. Save/Share: Keep your artistic creation or share it

🎨 Available Styles

The app includes various artistic styles inspired by:

• Van Gogh: Starry night and post-impressionist styles
• Picasso: Cubist and abstract interpretations
• Abstract Art: Modern artistic transformations
• Classical Paintings: Various historical art movements

⚡ Performance Metrics

• Inference Time: 4-5 seconds (average)
• Model Size: Optimized for mobile deployment
• Memory Usage: Efficient resource management
• Compatibility: Tested on Android 6.0+

🔬 Technical Deep Dive

Model Architecture

The app implements a custom Fast Neural Style Transfer architecture, redesigned from the original paper with significant optimizations for mobile deployment. Through architectural innovations and careful network design, the model achieves high-quality style transfer with only 210,000 parameters - a dramatic reduction from typical implementations while maintaining visual quality.

Mobile Optimization Strategy

• Custom Architecture Design: Redesigned Fast Style Transfer architecture from scratch
• Parameter Reduction: Optimized network to only 210,000 parameters through architectural innovations
• Per-Style Training: Individual model training from scratch for each artistic style
• Memory Management: Efficient buffer handling for mobile constraints

Android Integration

• Custom TensorFlow C_API Implementation: Direct native integration bypassing TensorFlow Lite limitations
• Dynamic Tensor Handling: Support for variable input dimensions without predefined sizes
• Zero Network Dependencies: Complete offline functionality - no server communication
• Privacy-Focused Architecture: All processing stays on device, no data transmission
• Asynchronous Processing: Non-blocking UI operations with real-time feedback
• Optimized Memory Management: Efficient resource handling for mobile constraints

🏗️ Build Instructions

Note: Source code is not publicly available. This is a demo application.

The app was built using:

1. Custom TensorFlow model training pipeline
2. Android Studio with NDK configuration
3. TensorFlow Lite model conversion and optimization
4. Custom native library compilation
5. JNI integration and testing

📊 Project Stats

• Development Year: 2021
• Model Training: Custom dataset and style images
• Testing: Multiple Android devices and versions
• Performance: Optimized for production use

🤝 Contributing

This is a demonstration project. For questions or collaboration opportunities, please reach out through GitHub issues.

📝 License

This project is a portfolio demonstration. Please contact for licensing information.

📧 Contact

Constantin Shafranski
Data Scientist | AI & ML Expert
LinkedIn | GitHub

---

Built with passion for mobile AI and artistic creativity 🎨📱