Progressive Growing of GANs

Module: Generative AI | Difficulty: Advanced

Progressive Growing

Fade-in

where increases from 0 to 1 during training.

StyleGAN2 Improvements

• Removed progressive growing
• Used style-based generator
• Added path length regularization

class ToRGB(nn.Module):
    def __init__(self, in_ch, out_ch=3):
        super().__init__()
        self.conv = nn.Conv2d(in_ch, out_ch, 1)
    def forward(self, x, alpha=1.0):
        return self.conv(x)

class FadeIn(nn.Module):
    def __init__(self, from_layer, to_layer):
        super().__init__()
        self.from_layer = from_layer
        self.to_layer = to_layer
    def forward(self, x, alpha):
        return (1-alpha)*self.from_layer(x) + alpha*self.to_layer(x)

Research Insight: Progressive growing was crucial for early GANs but became unnecessary with better regularization (spectral norm, path length reg). StyleGAN2 shows that architectural improvements can replace training tricks.