Generative AI

VAE & Simple GAN for Signature Generation

Generating realistic synthetic signatures with a VAE and a GAN.

2024

OVERVIEW

The Project

An implementation of a Variational Autoencoder (VAE) and a Generative Adversarial Network (GAN) to generate realistic fake signatures. The VAE learns latent representations by encoding and reconstructing signatures; the GAN's generator creates new signatures from noise while a discriminator judges realism.

Objectives

Learn meaningful latent representations of signatures with a VAE.

Generate convincing synthetic signatures with a GAN.

Improve robustness on a limited dataset via augmentation.

Tools & Technologies

PyTorchVAE (encoder/decoder)GAN (generator/discriminator)

FIGURES

Charts & Graphs

METHODOLOGY

The Approach

1

Augment the dataset (rotation, horizontal flip, Gaussian noise) for diversity.

2

Train the VAE with combined reconstruction + KL-divergence loss.

3

Train the GAN with binary cross-entropy adversarial loss.

4

Evaluate via reconstruction loss and qualitative comparison to real signatures.

OUTCOME

Results & Learnings

VAE achieved low reconstruction loss; GAN produced high-quality, realistic signatures.

Generated samples retained distinctive features of the originals.

Key Learnings

VAE structural understanding + GAN realism are complementary.
Augmentation meaningfully enriches small training sets.

Related Case Studies