This Artwork Doesn’t Exist

This repository contains the implementation and deployment of a Deep Convolutional Generative Adversarial Network (DCGAN) that generates novel artworks using a curated subset of the WikiArt dataset. Developed as part of Assignment 2 for the Generative AI & Applications course at SEECS, NUST by Basharat Ali, this project explores the intersection of technology and creativity.

Project Overview

• Objective:
  Train a DCGAN to learn artistic features from a diverse collection of artworks and generate original images.

• Dataset:
  A curated subset of 61,000 images from the WikiArt dataset was used. The images were preprocessed by:

  • Resizing to 64×64 pixels
  • Normalizing pixel values to a range of [-1, 1]

• Model Training:
  The DCGAN was trained using the following parameters:

  • Latent Dimension: 100
  • Batch Size: 1024
  • Number of Epochs: 150
  • Image Size: 64×64
  • Learning Rate: 0.0002
    Training was performed on Kaggle’s T4x2 GPUs.

Deployment

The trained model is deployed on Huggingface Spaces, where an interactive Gradio interface allows users to generate and explore artwork on the fly. The deployment code is available in the HF_App folder.

Repository Structure

• HF_App: Contains the code for model inference and the Gradio interface used for deploying the application on Huggingface Spaces.
• Training: Training and Inference code is in Training folder.

How to Use

1. Clone the repository:

   git clone https://github.com/BasharatWali/Wiki_ArtGAN.git