🎨 StyleForge AI

StyleForge AI is an AI-powered Neural Style Transfer web application built with Flask and PyTorch. It uses Adaptive Instance Normalization (AdaIN) to blend the artistic style of one image with the structural content of another, producing high-quality stylized artwork in just a few seconds.

The application provides an intuitive web interface where users can upload a content image and a style reference image, adjust the style strength, and generate a unique artistic output.

---

✨ Features

• 🖼️ Upload content and style images
• 🎨 AI-powered Neural Style Transfer using AdaIN
• 🎛️ Adjustable style strength slider
• ⚡ Fast inference with PyTorch
• 🖥️ Interactive Flask web interface
• 📥 Download generated stylized images
• 🖼️ Built-in example gallery
• 📱 Responsive modern UI
• 🧠 VGG Encoder + Trained Decoder architecture

---

🚀 Demo

Input

• Content Image
• Style Reference

↓

AI Processing

• VGG Encoder
• Adaptive Instance Normalization (AdaIN)
• Decoder Network

↓

Output

• Stylized AI Artwork

---

🛠️ Tech Stack

Backend

• Python
• Flask
• PyTorch
• Torchvision

Frontend

• HTML5
• CSS3
• Bootstrap 5
• JavaScript

Libraries

• Pillow
• Flask-WTF
• WTForms

---

📂 Project Structure

StyleForge-AI/
│
├── README.md
├── requirements.txt
├── Procfile.txt
├── code.ipynb
│
├── Demo_IO_Images/
│   ├── i-p/
│   └── o-p/
│
└── NST_Code/
    ├── app.py
    ├── train.py
    ├── vgg_normalised.pth
    ├── adain_algo.png
    │
    ├── content_data/
    ├── style_data/
    ├── examples/
    ├── static/
    │   └── uploads/
    │
    ├── templates/
    │   └── index.html
    │
    ├── experiment/
    │   └── final_exp/
    │       ├── decoder_final.pth
    │       └── sample_iter_*.png
    │
    └── utils/
        ├── models.py
        └── utils.py

---

⚙️ Installation

1️⃣ Clone Repository

git clone https://github.com/yourusername/StyleForge-AI.git

cd StyleForge-AI

---

2️⃣ Create Virtual Environment

Windows

python -m venv .venv

.\.venv\Scripts\activate

If activation is blocked:

Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass

.\.venv\Scripts\activate

---

3️⃣ Install Dependencies

pip install --upgrade pip

pip install -r requirements.txt

---

▶️ Run The Application

python NST_Code/app.py

Open your browser:

http://localhost:5000

---

Deploy On Render

This project does not commit large model files to GitHub. For Render deployment, upload these two local files to a public downloadable location such as Google Drive, Hugging Face, or GitHub Releases:

• NST_Code/vgg_normalised.pth
• NST_Code/experiment/final_exp/decoder_final.pth

In Render, create a Web Service or use the included render.yaml blueprint, then add these environment variables:

VGG_MODEL_URL=<public download URL for vgg_normalised.pth>
DECODER_MODEL_URL=<public download URL for decoder_final.pth>

Render start command:

python download_models.py && gunicorn --chdir NST_Code app:app

---

🎯 How It Works

Step 1

Upload a Content Image

↓

Step 2

Upload a Style Image

↓

Step 3

Adjust the Style Strength

↓

Step 4

Click Transfer Style

↓

Step 5

Download the generated artwork

---

🧠 Model Pipeline

Content Image
       │
       ▼
   VGG Encoder
       │
       ▼
Content Features

Style Image
       │
       ▼
   VGG Encoder
       │
       ▼
 Style Features

       │
       ▼
Adaptive Instance Normalization (AdaIN)

       │
       ▼
 Stylized Features

       │
       ▼
 Decoder Network

       │
       ▼
 Generated Artwork

---

🧩 Core AdaIN Function

adaptive_instance_normalization(content_feat, style_feat)

AdaIN transfers the statistical properties (mean and standard deviation) of the style features to the content features, enabling artistic style transfer while preserving the original image structure.

---

🏋️ Train the Model (Optional)

Training script:

cd NST_Code

python train.py \
--content_dir content_data \
--style_dir style_data \
--vgg vgg_normalised.pth \
--experiment my_exp \
--epochs 10

Note: GPU is recommended for training.

---

📸 Screenshots

Home Page

---

Generated Result

---

📁 Important Files

File	Description
app.py	Flask routes and inference pipeline
train.py	Model training script
models.py	VGG Encoder and Decoder architecture
utils.py	AdaIN implementation and utilities
index.html	Frontend interface
requirements.txt	Project dependencies

---

⚠️ Common Issues

Python not recognized

Install Python and enable

Add Python to PATH

Verify installation

python --version

---

Model file not found

Ensure the following files exist:

NST_Code/vgg_normalised.pth

NST_Code/experiment/final_exp/decoder_final.pth

---

Virtual Environment Error

Set-ExecutionPolicy -Scope Process -ExecutionPolicy Bypass

---

PyTorch Version Issue

pip install --upgrade -r requirements.txt

---

📌 Notes

• Supports JPG, JPEG, and PNG
• Generated images are stored inside:

NST_Code/static/uploads/

• First inference may take a few seconds because model weights are loaded.
• CUDA GPU is recommended for faster inference.

---

🔮 Future Improvements

• 🎭 Multiple built-in artistic styles
• ⚡ Real-time style preview
• 📱 Drag & Drop image upload
• 🖼️ Batch image stylization
• ☁️ Cloud deployment
• 📊 Image comparison slider
• 🧠 More advanced style transfer models

---

🙌 Acknowledgements

This project is inspired by the research on Arbitrary Style Transfer using Adaptive Instance Normalization (AdaIN) and is built using PyTorch, Flask, and VGG-based feature extraction.

---

👨‍💻 Author

Sumit Jaiswal

💻 Passionate about AI, Machine Learning, Computer Vision, and Full-Stack Development.

---

⭐ If you found this project useful, consider giving it a Star!