GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium

under the condition that critic learns faster than the actor

• separate learning rates
  • a(n) = o(b(n))
• the distriminator must learn new patterns before transfor to generator
• Adam: escape local minima that leads to mode collapse
• introducing FID: consistent with increasing disturbances and human judgement