🎨 TensorFlow Neural Style Transfer

A modular, production-ready implementation of Neural Style Transfer and image generation using TensorFlow 2 / Keras. Blend the content of any photo with the artistic style of any painting — in minutes.

Based on the seminal paper: "A Neural Algorithm of Artistic Style" — Gatys et al. (2015)

---

✨ Features

• 🖼️ Neural Style Transfer — fuse content & style images using VGG19 deep features
• ⚡ Fast Style Transfer — train a feed-forward network for real-time stylization
• 🧱 Modular Keras architecture — clean custom layers, loss functions, and training loops
• ⚙️ YAML config — full experiment control without touching code
• 📊 TensorBoard integration — live loss curves and image previews
• 💾 Checkpoint & resume — never lose a long stylization run
• 🖥️ CLI + Python API — use as a script or import as a library
• 🧪 Full test suite — pytest-ready unit tests

---

📁 Project Structure

tf-neural-style/
├── configs/
│   └── default.yaml          # Experiment configuration
├── src/
│   ├── model.py              # VGG19 feature extractor + Fast Style Net
│   ├── losses.py             # Content, style (Gram matrix), total variation loss
│   ├── stylizer.py           # Optimization-based style transfer engine
│   ├── fast_stylizer.py      # Feed-forward fast style transfer trainer
│   ├── preprocessing.py      # Image loading, resizing, normalization
│   └── utils.py              # Config, logging, callbacks, helpers
├── scripts/
│   ├── stylize.py            # Run optimization-based style transfer
│   ├── train_fast.py         # Train a fast style network
│   └── apply_fast.py         # Apply a trained fast style network
├── tests/
│   └── test_losses.py        # Unit tests for loss functions & model shapes
├── requirements.txt
└── README.md

---

🚀 Quick Start

1. Install

pip install -r requirements.txt

2. Run Style Transfer (optimization-based)

python scripts/stylize.py \
  --content assets/content.jpg \
  --style   assets/style.jpg \
  --output  outputs/result.jpg

3. Train a Fast Style Network

python scripts/train_fast.py \
  --style assets/starry_night.jpg \
  --dataset path/to/coco/train2017/

4. Apply Fast Style Network

python scripts/apply_fast.py \
  --checkpoint outputs/fast_model/ \
  --image path/to/photo.jpg \
  --output outputs/styled.jpg

---

🎛️ Configuration

Edit configs/default.yaml to control everything:

style_transfer:
  image_size: 512
  iterations: 1000
  content_weight: 1.0e4
  style_weight: 1.0e-2
  tv_weight: 30.0

---

📊 How It Works

Content Image ──► VGG19 ──► Content Features ──► Content Loss ──┐
                                                                  ├──► Total Loss ──► Adam ──► Stylized Image
Style Image ───► VGG19 ──► Gram Matrices ────► Style Loss ──────┘
                                               TV Regularization ┘

The algorithm iteratively updates pixel values of a generated image to minimize:

• Content loss — preserve high-level structure from content image
• Style loss — match Gram matrix statistics from style image
• Total variation loss — encourage spatial smoothness

---

🔧 Models

Mode	Speed	Quality	Use Case
Optimization (VGG19)	~5 min/image	⭐⭐⭐⭐⭐	Best quality, slow
Fast Style Net	~0.01s/image	⭐⭐⭐⭐	Real-time, per-style training

---

🧪 Tests

pytest tests/ -v

---

📄 License

MIT — free to use and modify.

---

📚 References

• Gatys et al. (2015) — A Neural Algorithm of Artistic Style
• Johnson et al. (2016) — Perceptual Losses for Real-Time Style Transfer