The main idea of BEGAN is to use an autoencoder on the discriminator, which will have its own loss that measures how well the autoencoder reconstructed some image (generator or real data):

Some advantages of BEGAN are as listed:

• High-resolution (128x128) face generation (2017 state of the art) .
• Offers a way to measure convergence .
• Good results even without batch-norm and dropout .
• Hyperparameter to control the generation diversity versus quality. More quality also means more mode collapse.
• Having two separate optimizers are not required .

Here's an example of the quality of images that a BEGAN can produce when tasked with generating human faces: