1. GAN Improvements

1-1. Stability

Mode collapse

• A single minibatch can have high or low standard deviation

Mini batch std $\uparrow$ → diversity $\uparrow$

Mini batch std $\downarrow$ → diversity $\downarrow$

Use standard deviation in batch to encourage diversity

Gradient penalty

• Improve stability by enforcing 1-Lipschitz continuity
• E.g WGAN-GP, Spectral Normalization

Averaging weights

• Use Moving average for smoother results

Progressive Growing

• Progressive growing gradually trains GAN at increasing resolutions

1-2. Capacity & Diversity

• Larger models and higher resolution images

• Increasing variety on generated images