Neural Style Transfer: Transferring Painting Styles

This repository contains an implementation of Neural Style Transfer, allowing users to apply the stylistic elements of famous paintings to photographic images while retaining the original color palette of the content images. The project is based on Gatys et al.'s algorithm with key modifications to improve color retention and flexibility.

Project Overview

This project uses Convolutional Neural Networks (CNNs)(VGG architecture) to merge the content of one image with the style of another. We extend traditional methods by introducing a pre-transfer color alignment process to retain the color integrity of the content image, enabling robust and artistic image transformations.

Key Features:

• Style Transfer: Transforms a content image into a new image by applying the style of a famous painting.
• New method of Color Retention: Preserves the original color palette of the content image using a color alignment process trough a linear transformation.
• Customizable Style and Content Weights: Fine-tune the balance between style and content using adjustable hyperparameters.

Final Report:

Final Report Link

Technologies:

• PyTorch 2.3.0 (for neural network implementation)
• VGGNet-19 (pre-trained on ImageNet)
• Python 3.10
• NumPy 1.26.0
• PyTorch 2.3.0
• PyTorch-cuda 11.8
• Pillow 10.4.0
• TorchVision 0.18.0
• OpenCV 4.8.1

Contributors

• Giovanni Benedetti da Rosa(me)
• Paulo Roberto de Moura Júnior
• Juan Esteban Rios Gallego
• Cristian Alejandro Chávez Becerra

Original repo