Training a generator to make a detailed face asks it to solve several problems at once. It needs a plausible overall arrangement, coherent features and fine texture. If the large-scale structure is wrong, beautifully detailed pores don't help much.
Progressive Growing of GANs, by Tero Karras and colleagues, changes the order of that problem. Both the generator and discriminator begin at a low resolution. New layers are introduced as training proceeds, allowing the system to work at progressively higher resolutions. The released code makes the training schedule a central part of the implementation.
The discriminator matters because a GAN is a coupled system. The generator tries to produce plausible samples, while the discriminator learns to distinguish generated images from training examples. Growing only one side would change the challenge abruptly. Growing both sides lets their responsibilities expand together.
The transition isn't simply a hard switch from a small image to a large one. Newly added layers are faded into the computation. During that period, the old lower-resolution path and the new higher-resolution path both contribute. This gives the existing system a chance to adjust instead of suddenly receiving an entirely different problem.
I find the idea useful as a training curriculum expressed in architecture. Early work concentrates on coarse structure because that is all the current representation can express. Later work has the capacity to refine details. The schedule does not tell the network where an eye belongs, but it controls the scale at which mistakes can be represented and challenged.
There is an important distinction between a sharp output and a good model. A generator can produce convincing individual images while covering only a narrow part of the training distribution. Diversity matters. So do repeated structures, strange backgrounds and failures that become visible only when looking across many samples.
A small evaluation exercise would therefore include several views of the output: individual images at full resolution, contact sheets that make repetition easier to notice, and transitions through the input space. Each reveals a different failure. Choosing only the best-looking face is closer to selecting an illustration than evaluating a generator.
Progressive training also doesn't remove the influence of the dataset. If examples are framed consistently, the model has an easier problem than if pose, crop and background vary wildly. Data preparation remains a design decision, not housekeeping before the interesting part.
What attracts me here is the modesty of the central mechanism. Change when the network has to learn each scale of structure, and the behavior of the whole training process changes. Sometimes a better route to detail begins by temporarily refusing to represent it.