Real-time artistic style transfer web app powered by Adaptive Instance Normalization (AdaIN). Upload any content image and a style painting — the neural network transfers the artistic style instantly.

🔗 Live Demo: tejas-m-stylecast.hf.space

StyleCast uses a three-step pipeline based on the AdaIN algorithm (Huang & Belongie, ICCV 2017):

1. Encode — Both content and style images pass through a pre-trained VGG-19 encoder, extracting deep feature representations.
2. AdaIN Transfer — Content features are normalized to match the mean and variance of the style features, aligning their statistical distributions in feature space.
3. Decode — A trained decoder network converts the transformed features back to pixel space, producing the final stylized image.

The alpha slider (0–1) controls how much style is applied — 0 preserves the original content, 1 applies full style transfer.

Content	Style	Output

The decoder was trained from scratch in two phases:

• Content dataset: ~40,000 images from MS-COCO
• Style dataset: ~8,000 paintings from WikiArt
• Encoder: Pre-trained VGG-19 (frozen weights)
• Decoder: Trained from scratch
• Loss: Content loss (MSE on relu4_1 features) + Style loss (MSE on mean/std across relu1_1 to relu4_1)
• Optimizer: Adam
• Hardware: NVIDIA RTX 5050

• PyTorch — Model training and inference
• Flask — Web backend
• VGG-19 — Pre-trained encoder
• Gunicorn — Production server
• Docker — Containerized deployment
• HuggingFace Spaces — Hosting

StyleCast/
├── app.py                  # Flask web app
├── train.py                # Training script
├── decoder_final.pth       # Trained decoder weights
├── vgg_normalised.pth      # Pre-trained VGG-19 encoder
├── Dockerfile              # Docker config for deployment
├── Procfile                # Process config
├── requirements.txt        # Python dependencies
├── templates/
│   └── index.html          # Frontend
├── examples/               # Example images
├── static/uploads/         # User uploads
├── content_data/           # Sample content images
├── style_data/             # Sample style images
└── utils/
    ├── models.py           # VGG Encoder & Decoder architectures
    └── utils.py            # AdaIN, dataset, transforms

# Clone the repo
git clone https://github.com/iam-teju/StyleCast.git
cd StyleCast

# Install dependencies
pip install -r requirements.txt

# Run the app
python app.py

Open http://localhost:5000 in your browser.

• Huang, X., & Belongie, S. (2017). Arbitrary Style Transfer in Real-time with Adaptive Instance Normalization. ICCV 2017.

Built by Tejas M · LinkedIn · GitHub