What does "progressive growing" in GAN refer to?

The term "progressive growing" in the context of Generative Adversarial Networks (GANs) specifically refers to a strategy where the model starts training at a low resolution and gradually increases the output resolution over time. This approach allows the GAN to learn basic features of the data at first, effectively reducing the complexity of the initial training phase. As the training progresses, the model adds more layers to capture finer details, leading to better quality outputs.

This method is particularly beneficial in managing computational resources and helps stabilize the training process by preventing potential issues that can arise from starting with high-resolution images, such as mode collapse or difficulty in convergence. By focusing on features progressively and expanding the resolution step by step, the GAN can produce high-quality images without overwhelming the model at the outset.