How can I tell if my GAN discriminator is the bottleneck behind fake outputs?

[Ella77]

I’ve been trying to understand the role of the discriminator in my GAN setup, but I’m hitting a wall. My generator’s output still looks obviously fake, and I can’t tell if the issue is with my discriminator’s architecture or if my training loop’s feedback is just too weak.

[Ryan_J]

I spent weeks really staring at the discriminator in my GAN. At first the generator stayed painfully obvious as fake, and I blamed the generator. Then I watched the losses and saw D barely moved while G kept outputting near misses. I swapped in a lighter architecture and added label smoothing, and for a moment the generator looked less obviously fake, but it never fully clicked.

[Jerry_J]

I tried bigger batches, different learning rates, even a different optimizer, and nothing jumped out. It felt like the feedback loop wasn't giving the generator enough signal, or maybe the data pipeline was quietly starving it.

[Lily_M]

Could the real problem be the training signal itself rather than the architecture?

[VioletA]

I sometimes went off topic and started logging every generated image just to see patterns, and a handful of samples looked decent after a late night run, only to vanish the next morning. It made me wonder if the issue is consistency in training rather than any single component.