Generative Models - GANs v.s. VAEs v.s. Diffusion Models

comparision of GANs, VAEs and Diffusion Models

From: https://pub.towardsai.net/diffusion-models-vs-gans-vs-vaes-comparison-of-deep-generative-models-67ab93e0d9ae

GANs

• GAN = Generator + Discriminator
• Training loss: adversarial loss. The generator aims to “fool” a discriminator.
• High-fidelity results. The discriminator cannot distinguish between the fake and real samples.
• Low-diversity results (mode collapse): When the discriminator has over-trained or catastrophic forgetting happens, the generator might be happy enough to produce a small part of data diversity.
• Hard to train. It can difficult to determine when your network converged.

VAEs

• VAE = Encoder + Decoder
• Training by maximizing log-likelihood.