DCGAN: Deep Convolutional GAN Architecture
You'll learn: How DCGAN revolutionized GANs by introducing a stable convolutional architecture that produces high-quality images.
The Problem DCGAN Solved
Early GANs were notoriously unstable and used fully-connected layers, making them poor at generating images. DCGAN (Deep Convolutional GAN) introduced a specific architectural design that brought stability and quality to image generation.
Core Design Principles
DCGAN replaces fully-connected layers with convolutional operations throughout both networks:
Generator Architecture:
- Starts with a random noise vector
- Uses transposed convolutions (also called "deconvolutions") to progressively upsample
- Think of transposed convolutions as "reverse convolutions" that expand spatial dimensions
- Example path: 100-dim noise → 4×4 feature maps → 8×8 → 16×16 → 32×32 → 64×64 image
Discriminator Architecture:
- Takes an image as input
- Uses strided convolutions (stride > 1) to progressively downsample
- No pooling layers—downsampling happens through convolution stride alone
- Example path: 64×64 image → 32×32 → 16×16 → 8×8 → 4×4 → real/fake classification
Why This Works
Transposed convolutions learn how to upsample intelligently, while strided convolutions learn what features matter for discrimination. This creates a cleaner gradient flow compared to fully-connected layers, reducing training instability.
The architecture also uses batch normalization (except generator output and discriminator input), ReLU in the generator, and LeakyReLU in the discriminator—all contributing to stable training.
Key Takeaway: DCGAN established the canonical GAN architecture using transposed convolutions in the generator to upsample and strided convolutions in the discriminator to downsample, creating stable training and high-quality image generation.