Generative Models for Signature Synthesis and Face Translation: A Comparative Study of VAE, GAN, Conditional GAN, and CycleGAN

<title>Abstract</title> <p>This paper presents a comparative implementation of four generative modeling approaches: a Variational Autoencoder (VAE), a standard Generative Adversarial Network (GAN), a Conditional GAN (cGAN), and CycleGAN. The experiments address three image-generation settings: signature synthesis, cat-and-dog image generation using a CIFAR-10 subset, and bidirectional translation between facial sketches and face photographs. The study describes dataset preparation, preprocessing, model design, training objectives, and evaluation procedures. In the reported experiments, the VAE obtained a test reconstruction loss of 0.015, the signature GAN discriminator achieved 85% classification accuracy, the custom CIFAR-10 GAN obtained a similarity score of 0.72, and the cGAN achieved an average pixel-similarity score of 0.80. For CycleGAN, the cycle-consistency loss stabilized after approximately 50 epochs and the generated translations became visually more coherent. The experiments highlight the strengths of latent-variable, adversarial, conditional, and cycle-consistent learning, while also illustrating common challenges such as adversarial instability, sensitivity to hyperparameters, and the limitations of pixel-level evaluation for perceptual image quality.</p>