A team trains a GAN on photographs of flowers. They log how often the discriminator is right when shown a batch that is half real photographs and half generated ones, so pure guessing scores :
| epoch | discriminator accuracy |
|---|---|
They log the discriminator's raw output too. For the whole run it returns a "probability real" between and for every example it is shown — generated ones and genuine photographs alike.
The samples at epoch are formless grey smudges. They are no better than the samples at epoch .
A teammate reads the table and calls the run finished: "A judge reduced to coin-flipping is the definition of success here — that is the target, and we hit it. The pictures just need more epochs to sharpen."
What has actually happened?
Select all that apply.