Progressive Growing of GANs
What You'll Learn
You'll learn how Progressive Growing of GANs (ProGAN) trains generators and discriminators by starting small and gradually scaling up to high-resolution images for better stability and quality.
The Core Idea
Imagine learning to paint: you'd start with rough sketches before adding fine details. Progressive Growing of GANs follows the same principle. Instead of training a GAN to generate high-resolution images (like 1024×1024) from the start—which is unstable and difficult—ProGAN begins at very low resolution (4×4 or 8×8) and gradually adds layers to both the generator and discriminator to increase resolution step-by-step.
How It Works
The training progresses through stages:
- Start Low: Train generator and discriminator at 4×4 resolution until stable
- Add Layers Smoothly: Introduce new layers for the next resolution (8×8), gradually "fading in" their contribution
- Stabilize: Train at this new resolution until convergence
- Repeat: Continue doubling resolution (16×16 → 32×32 → 64×64 → 512×512 → 1024×1024)
During the fade-in phase, the network smoothly transitions between resolutions using a weighted blend. This prevents the sudden introduction of new layers from destabilizing training—a common problem with standard GANs at high resolutions.
Why It Matters
This approach solves several GAN training challenges you've learned about:
- Reduces mode collapse by ensuring stable training at each resolution
- Prevents vanishing gradients by starting simple and adding complexity gradually
- Improves convergence since the network learns coarse structure first, then refines details
- Enables high-resolution generation that was previously impossible
Key Takeaway: Progressive Growing trains GANs by incrementally adding layers that double resolution at each stage, starting from low resolution and smoothly transitioning upward—dramatically improving stability and enabling generation of high-quality, high-resolution images.