# Art Style Transfer

This project implements **artistic style transfer** using PyTorch. It blends the artistic style of an image (e.g., a painting) into the content of another image (e.g., a photograph). The implementation uses a pretrained VGG19 model for feature extraction and computes content and style losses to optimize the generated image.

---

## 📂 Project Structure

Art_Style/ ├── artwork/ # Folder for style artwork images ├── content/ # Folder for content images ├── style/ # Folder for style images ├── generated/ # Folder for saving generated images (notebook) ├── generated_images/ # Folder for saving generated images (Streamlit app) ├── app.py # Streamlit web application ├── Art_Style_Transfer_Model.ipynb # Jupyter Notebook for experimentation ├── README.md # Project documentation └── requirements.txt # Python dependencies

---

## ✨ Features

- **Blend Styles**: Merges artistic styles into content images using neural networks.
- **Pretrained Model**: Uses a VGG19 model pretrained on ImageNet.
- **Intermediate Visualization**: Displays the generated image every 100 steps during optimization.
- **Output Storage**: Saves the final result in the `generated/` or `generated_images/` folder.

---

## 🔬 Methodology

This project implements a neural style transfer approach based on VGG19 for separating content from style information:

- **Deep Feature Extraction**: Content features are extracted from deeper convolutional layers of the VGG19 network, capturing high-level semantic information while preserving structural elements.

- **Style Modeling**: Artistic style is modeled using **Gram matrices** computed across multiple convolutional layers. Gram matrices capture the statistical correlations between feature maps, effectively representing texture and artistic patterns without preserving spatial layout.

- **Loss Function Optimization**: The stylized image is optimized through a composite loss function that balances:
  - **Content Loss**: Ensures the generated image preserves the structural content of the original
  - **Style Loss**: Encourages artistic expression by matching Gram matrices from the style image
  - **Total Variation Loss** (optional): Promotes spatial smoothness and reduces artifacts

- **Evaluation Metrics**: Experimental validation includes:
  - **PSNR** (Peak Signal-to-Noise Ratio): Measures reconstruction quality
  - **SSIM** (Structural Similarity Index): Assesses perceptual similarity
  - **TV-Loss Ablation Studies**: Demonstrates the impact of total variation regularization

The method achieves successful stylization while maintaining structural integrity and visual quality.

---

## 🔧 Installation

1. **Clone the repository:**
   ```bash
   git clone https://github.com/your_username/Art_Style.git
   cd Art_Style

2. Install dependencies:

   pip install -r requirements.txt

3. Verify PyTorch installation (optional): Ensure that PyTorch is installed with CUDA support if you want to use a GPU.

   python -c "import torch; print(torch.cuda.is_available())"

---

🚀 How to Use

Option 1: Streamlit Web App

1. Run the Streamlit application:

   streamlit run app.py

2. Upload images:

   • Upload your content image through the web interface
   • Upload your style image
   • Adjust hyperparameters (alpha, beta, steps, TV loss weight)
   • Click "Start Style Transfer"

3. View results:

   • Generated images are saved in generated_images/ folder
   • View loss plots, metrics (PSNR, SSIM), and intermediate images
   • Run ablation experiments (α/β sweep, TV loss comparison)

Option 2: Jupyter Notebook

1. Prepare Images:

   • Place your content images in the content/ folder.
   • Place your style images in the style/ folder.

2. Modify File Paths: Update the content_file and style_file paths in the notebook:

   content_file = "content/dog.png"
   style_file = "style/ab4.jpeg"

3. Run the Notebook: Open and run Art_Style_Transfer_Model.ipynb in Jupyter.

4. View Results: Generated images will be saved in the generated/ folder (or path specified in the notebook).

---

📊 Example Outputs

Content Image	Style Image	Generated Image

---

⚙️ Hyperparameters

You can tweak these parameters in app.py (via UI sliders) or Art_Style_Transfer_Model.ipynb to experiment with different results:

• alpha: Weight for content loss (default: 1.0). Higher values preserve more content structure.
• beta: Weight for style loss (default: 0.1). Higher values apply stronger artistic style.
• total_steps: Number of optimization steps (default: 3000 for notebook, 500 for app). More steps = better results but longer processing.
• lr: Learning rate for the optimizer (default: 0.0005). Controls the step size during optimization.
• tv_weight: Total variation loss weight (default: 0.0 or 1e-6). Promotes smoothness and reduces artifacts.

---

📜 Dependencies

The project requires the following libraries (listed in requirements.txt):

• torch (PyTorch) - Deep learning framework
• torchvision - Pre-trained models and image transforms
• numpy - Numerical computations
• Pillow - Image processing
• tqdm - Progress bars
• scikit-image - Image metrics (PSNR, SSIM)
• pandas - Data manipulation and analysis
• streamlit - Web application interface (for app.py)
• matplotlib - Plotting and visualization

Install them with:

pip install -r requirements.txt

---