Debugging a Restoration GAN That Won't Sharpen

You're training an ESRGAN-style restoration generator. Training is stable (no mode collapse, no vanishing-gradient symptoms as covered in the GAN foundation subject), and the discriminator's loss looks reasonable throughout training, but the generator's output stays noticeably softer than reference ESRGAN/Real-ESRGAN outputs at a comparable training budget. A teammate suggests the discriminator is "too weak" and proposes increasing its capacity.

1. Before changing discriminator capacity, what is the first thing you would check about the discriminator's formulation, and why?
2. Explain what a relativistic discriminator changes relative to a standard (non-relativistic) discriminator, and why that change specifically helps generate sharper texture rather than just "training more stably."
3. Is the teammate's "increase discriminator capacity" proposal a reasonable next step if the discriminator turns out to already be relativistic? Explain your reasoning.