Deep Convolutional GAN (DCGAN)

← Back to Synthetic Data Generation

Generative Adversarial Networks (GANs) [14] train a generator network \(G\) and a discriminator network \(D\) in an adversarial game:

\[\min_G \max_D \; \mathbb{E}_{x \sim p_{\text{data}}}[\log D(x)] + \mathbb{E}_{z \sim p_z}[\log(1 - D(G(z)))]\]

The DCGAN architecture [15] uses convolutional layers in both networks, enabling high-quality image synthesis. Once trained, the generator can sample arbitrarily many synthetic images by feeding random noise vectors \(z \sim p_z\).

Limitations: GAN training is notoriously unstable (mode collapse, vanishing gradients); the generated images may not be paired with segmentation masks without additional work (e.g., paired cGAN).

See also

• Stable Diffusion Image-to-Image — a pretrained, prompt-conditioned alternative that avoids training a generator from scratch

• Physics-Based Synthesis — a ground-truth-by-construction alternative that sidesteps the mask-pairing limitation

• Synthetic Data Generation — every other synthetic data strategy