Comprehensive Overview of GANs: History, Improvements, Applications, and Handwriting Style Transfer

Introduction

Since Ian Goodfellow introduced Generative Adversarial Networks (GAN) in 2014, many variants have emerged, improving loss functions, architectures, and applications such as style transfer and image‑to‑image translation.

Early GAN

Original GAN learns a mapping G(z) from a random noise distribution to a target data distribution by adversarial training of a generator G and a discriminator D. The training alternates between updating D to distinguish real from fake data and updating G to fool D.

Key equations and a reference implementation are shown.

Architecture

The first GAN used a multilayer perceptron for both generator and discriminator.

Loss Function

The minimax objective V(D,G) is a log‑likelihood for a binary classifier; in practice the discriminator uses binary cross‑entropy loss.

Major Improvements

DCGAN

Deep Convolutional GAN replaces MLPs with convolutional layers, uses batch normalization, ReLU (except tanh at the output), and fractional‑strided convolutions for up‑sampling.

WGAN

Wasserstein GAN substitutes the Jensen‑Shannon divergence with the Earth‑Mover (Wasserstein) distance, leading to more stable training and alleviating mode collapse and gradient vanishing.

Mode Collapse and Gradient Vanishing

Mode collapse occurs when the generator produces limited variety; it is caused by vanishing gradients in the original GAN loss. WGAN’s Wasserstein loss mitigates this problem.

Image‑to‑Image Translation Applications

GANs enable image‑to‑image translation and style transfer. Supervised methods include conditional GAN (cGAN) and pix2pix; unsupervised methods include CycleGAN, which introduces cycle‑consistency loss.

cGAN

Conditioning the generator on an additional vector allows control over generated attributes.

pix2pix

Uses paired images and a U‑Net generator with PatchGAN discriminator; combines adversarial loss with L1 loss.

CycleGAN

Learns mappings between two domains without paired data by enforcing that translating an image there and back yields the original.

Practical Handwriting Style Transfer (Laiye Technology)

Two projects are described:

BicycleGAN‑based English/number handwriting style transfer.

Zi2Zi‑based Chinese character style transfer.

Both use encoder‑decoder generators, multiple loss terms (adversarial, L1, KL, classification, TV), and address issues such as stroke loss.

Advantages and Disadvantages of GANs

Advantages: simple back‑propagation training, high sample quality, wide range of image applications, end‑to‑end training.

Disadvantages: unstable training, difficulty with discrete data, gradient‑vanishing and mode‑collapse issues.

Training Tips

Typical tricks focus on the traditional noise‑to‑image GAN; references are provided.

Application Scenarios

GANs are used for image synthesis, face generation, super‑resolution, inpainting, video prediction, 3D object generation, and many other visual tasks.

References

List of seminal papers and URLs covering GAN fundamentals, DCGAN, WGAN, conditional GAN, pix2pix, CycleGAN, BicycleGAN, Zi2Zi, and related tutorials.