PROGRESSIVE GROWING OF GANS FOR IMPROVED QUALITY, STABILITY, AND VARIATION

generative models

• autoregressive models: sharp images but slow, no latent representations
• VAE: easy to train but blurry results
• GAN: shapr but limited variations
  • gradients points to random directions when fake & real image not overlapping
  • higher res. makes discrimination easier
  • unhealthy competition when signal magnitute escalate in G and D

PROGRESSIVE GAN

• coarse to fine: (smoothly) add single layer once at a time
• learning an easy task at one stage
• batch std: each feature at (x, y), avg, concat
• normalize layer by c(from He. et al.), normalize feature vector after each layer
• similarity to Laplacian pyramid