Neural Style Transfer (Optimization-based)

This project implements classical Neural Style Transfer (Gatys et al.), where a generated image is directly optimized to preserve the content of one image and the style of another using a fixed pretrained VGG network.

Unlike fast style transfer approaches, this method does not train a new model.
The output image itself is optimized via gradient descent.

---

Method Overview

• Uses VGG19 pretrained on ImageNet as a fixed feature extractor
• Content is preserved using high-level feature representations
• Style is represented using Gram matrices
• The generated image is optimized to minimize:

$$ \mathcal{L} = \alpha \cdot \mathcal{L}_{content} + \beta \cdot \mathcal{L}_{style} $$

No model weights are trained or saved.
The final stylized image is the result of the optimization process.

---

Demo Results

Content Image

---

Style Image

Stylized Result

---

Style Image

Stylized Result

---

Style Image

Stylized Result

---

Project Structure

image-style-transfer/
├── notebooks/
│ └── 01_exploration.ipynb
├── src/
│ ├── utils.py
│ ├── model.py
│ ├── loss.py
│ ├── run_style_transfer.py
├── experiments/
│ └── exp_01/
│ └── config.yaml
│ └── notes.md
├── outputs/
│ └── images/
├── scripts/
│ └── run_exp.sh
├── requirements.txt
└── README.md

---

Installation

Requirements

• Python 3.9+
• PyTorch

Installing PyTorch may take several minutes.

Setup (CPU-only recommended)

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Customize hyperparam in config.yaml

Run

python3 src/run_style_transfer.py config.yaml