This repository contains the final group project for the Advanced Computer Vision course. Our project explores two core computer vision challenges applied to an art dataset: • Style Transfer – applying the style of a reference artwork (e.g., Cubism) to a target image • Fake Image Detection – classifying whether an image was generated by AI or created by a human artist

For each task, we develop and compare two models: a Champion model (best performing) and a Challenger model (alternative approach), with the goal of deploying the top models for real-time usage via a web app.

🔍 Project Overview

🎨 1. Style Transfer

We use deep learning to transfer artistic styles to input images.

Models:

• Champion: VGG-based Neural Style Transfer using a pretrained model and Cubism-style reference image
• Challenger: Custom CNN architecture trained from scratch

This task focuses on visual fidelity, stylization quality, and processing speed.

🖼️ 2. Fake Image Detection

We classify whether an artwork is AI-generated or human-made using two architectures and two dataset configurations (pure vs. hybrid).

Models:

• Champion: Pre-trained DINOv2 with a custom classification head (linear regression for logits)
• Challenger: EfficientNet

Evaluation includes accuracy, F1-score, ROC-AUC, and performance on a hybrid dataset.

⚙️ Methodology

• Preprocessing: Image normalization, resizing, label encoding
• Training: Implemented with PyTorch, early stopping and LR scheduling used
• Loss Functions:
  • Style Transfer: Content loss + Style loss
  • Fake Detection: Binary Cross-Entropy Loss
• Evaluation: Visual and quantitative metrics, ROC curves for fake detection
• Deployment: Streamlit app

🚀 Deployment

The final models were deployed via a Streamlit application with optional ngrok or localtunnel integration for live public demos. The web app supports:

• Uploading new images for style transfer
• Uploading test images to classify as real or AI-generated