superseded by denoising diffusion, which covers the data distribution instead of collapsing onto part of it and trains stably; GANs did not vanish, they became a loss term in few-step distillation
Learn more →
Why study this historical topic?
superseded by denoising diffusion, which covers the data distribution instead of collapsing onto part of it and trains stably; GANs did not vanish, they became a loss term in few-step distillation
Adversarial Training as the Default Route to Photorealistic Image Synthesis
students arrive believing image generation means training a GAN; naming that belief as the superseded default is what lets the diffusion half of the spine land
GANs in Machine Learning: Generative Adversarial Networks
Generative Adversarial Networks (GANs) represent a framework in unsupervised learning defined by a zero-sum game between two neural networks: a generator and a discriminator. The core mechanism relie…