GAN for Anime Face Generation

This project implements a Deep Convolutional GAN in PyTorch to generate anime-style face images. The model uses a convolutional generator and discriminator trained adversarially on the Anime Face Dataset, with generated samples and checkpoints saved throughout training.

Source code: Github

How the model works

A GAN is built from two competing neural networks that improve through adversarial training.

• The generator takes a random latent vector and learns to transform it into a realistic 64x64 anime face. At first, it produces noisy images, but over time it learns to create structure such as eyes, hair, and facial contours.
• The discriminator acts as a classifier that tries to distinguish real images from generated ones. It learns to assign high scores to authentic samples and low scores to synthetic ones.
• During training, the generator and discriminator are optimized against each other: the generator tries to fool the discriminator, while the discriminator tries to become better at spotting fakes. This competition drives both networks to improve over time.

What I built

• Trained a DCGAN architecture for 64x64 anime-face generation.
• Used a latent vector of size 128 with Adam optimizers and binary cross-entropy loss.
• Saved generated sample images, model weights, and a training video from the image sequence.

Project files

• gan.ipynb — main training notebook
• G.pth and D.pth — saved generator and discriminator weights
• generated/ — sample outputs produced during training
• anime_gans_training.avi — video of training progression

Notes

The notebook was designed for the Anime Face Dataset and can be run locally with PyTorch, torchvision, matplotlib, tqdm, numpy, pandas, and OpenCV.