Wasserstein GAN for Anime Face Generation

---

Brief Overview

In this project, I have learned and implemented a Wasserstein GAN with Gradient Penalty in PyTorch.

The aim of this work is to:

• Explore stable WGAN training using Wasserstein loss with Gradient Penalty.
• Display the generator and critic architectures using torchview
• Generate visually realistic anime faces from noise vectors of dim=100.
• Evaluate quality using Fréchet Inception Distance (FID).
• Provide an interactive demo website for trying the model with streamlit.

I developed the project end-to-end:

• Designed Generator & Critic architectures
• Built the training loop with visualization & checkpoints
• Added evaluation pipeline with FID scoring
• Deployed a Streamlit app for real-time face generation

Try it yourself here: Live

---

Output

Generated anime faces after training (FID ≈ 23.9):

---

Project Structure

GAN-Project/
│
├── README.md
├── requirements.txt
├── LICENSE
├── .gitignore
├── app.py
├── model.py
│
├── notebooks/
│   ├── WGAN-GP_trainer.ipynb
│   ├── WGAN-GP_evaluator.ipynb
│
├── results/
│   ├── generated (3).png
│   ├── generated (2).png
│   ├── generated (1).png
│   ├── generated.png
│   ├── generator_architecture.png
│   ├── critic_architecture.png
│
├── models/
│   ├── discriminator31.pth
│   ├── generator31.pth
│   ├── filter.pth

---

Installation

Clone the repository and install dependencies:

git clone https://github.com/PradumnS-001/GAN-Project
cd WGAN-AnimeFaces
pip install -r requirements.txt

---

Train Model

jupyter notebook notebooks/trainer.ipynb

Evaluate Model

jupyter notebook notebooks/evaluator.ipynb

Run Interactive App

streamlit run app/app.py

---

Dataset: Anime Faces Dataset (Kaggle) Link: https://www.kaggle.com/datasets/splcher/animefacedataset

Note: Dataset is NOT included in this repo. Please download separately.