Neural Style Transfer

This project implements Neural Style Transfer using PyTorch. Neural Style Transfer is a technique that takes a content image and a style image and generates a new image that combines the content of the first image with the style of the second.

Features

• Optimization-based style transfer: Apply style transfer by optimizing an image directly
• Fast neural style transfer: Use pre-trained models for instant stylization
• Web application: User-friendly interface for style transfer
• Multiple style transfer options: Control style weight, content weight, and iterations
• Batch processing: Process multiple images with one command
• Animation generation: Create GIFs showing the style transfer process

Requirements

• Python 3.7+
• Dependencies listed in requirements.txt

Installation

1. Clone this repository:

git clone https://github.com/yourusername/neural-style-transfer.git
cd neural-style-transfer

2. Install the required dependencies:

pip install -r requirements.txt

3. Download sample images (optional):

python download_samples.py

Usage

Command-line Style Transfer

Run the main style transfer script:

python style_transfer.py --content images/content/your_content_image.jpg --style images/style/your_style_image.jpg --output output/result.jpg

Parameters

• --content: Path to content image
• --style: Path to style image
• --output: Path for saving the output image
• --iterations: Number of optimization steps (default: 300)
• --style-weight: Weight for style loss (default: 1000000)
• --content-weight: Weight for content loss (default: 1)
• --lr: Learning rate (default: 0.01)
• --plot: Show a plot of the content, style, and output images

Demo with Progress Visualization

Run the demo script to see the style transfer process with visualization:

python demo.py --content images/content/your_content_image.jpg --style images/style/your_style_image.jpg --output-dir output/demo

Parameters

• --content: Path to content image
• --style: Path to style image
• --output-dir: Directory to save result images
• --iterations: Number of optimization steps (default: 300)
• --style-weight: Weight for style loss (default: 1000000)
• --content-weight: Weight for content loss (default: 1)
• --lr: Learning rate (default: 0.01)
• --save-interval: Interval for saving progress images (default: 50)
• --no-animation: Disable creation of GIF animation

Batch Processing

Process multiple images at once:

python batch_process.py --input-dir images/content --model models/starry_night.pth --output-dir output/batch

Parameters

• --input-dir: Directory containing images or a single image path
• --model: Path to a saved model for fast style transfer
• --output-dir: Directory to save stylized images
• --size: Size of the output images (default: 512)

Training Your Own Style Model

Train a fast neural style transfer model with your own style:

python train.py --dataset path/to/dataset --style-image images/style/your_style.jpg --checkpoint-dir checkpoints/your_style

Parameters

• --dataset: Path to training dataset (should contain content images)
• --style-image: Path to style image
• --checkpoint-dir: Directory to save checkpoints
• --epochs: Number of training epochs (default: 2)
• --batch-size: Batch size for training (default: 4)
• --content-weight: Weight for content loss (default: 1.0)
• --style-weight: Weight for style loss (default: 5.0)

Web Application

Run the Flask web app:

python app.py

Then open your browser and go to http://localhost:5000/

The web app allows you to:

• Upload content and style images
• Choose between optimization-based or fast style transfer methods
• Adjust style transfer parameters
• Download stylized images

Implementation Details

The implementation uses the VGG19 model pretrained on ImageNet to extract features from content and style images. Two main approaches are implemented:

1. Optimization-based style transfer: Directly optimizes a generated image to match the content features of the content image and the style features (Gram matrices) of the style image.

2. Fast neural style transfer: Uses a trained transformation network (TransformerNet) that can stylize images in a single forward pass, providing much faster results after training.

Examples

Example command:

python style_transfer.py --content images/content/landscape.jpg --style images/style/starry_night.jpg --output output/landscape_starry.jpg --iterations 500

Project Structure

• style_transfer.py: Main implementation of neural style transfer
• models.py: Neural network models for feature extraction and transformation
• utils.py: Utility functions for image processing and visualization
• demo.py: Script for demonstrating style transfer with visualization
• train.py: Script for training fast neural style transfer models
• stylize.py: Script for applying trained style models to images
• batch_process.py: Script for batch processing multiple images
• app.py: Flask web application for style transfer
• download_samples.py: Script to download sample content and style images
• templates/: HTML templates for the web application
• static/: Static files for the web application

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• The implementation is inspired by the paper "A Neural Algorithm of Artistic Style" by Gatys et al.
• Fast neural style transfer is based on "Perceptual Losses for Real-Time Style Transfer and Super-Resolution" by Johnson et al.