From traditional GANs to StyleGAN: architecture trace

What's actually being asked

Walk through the architectural evolution step by step — traditional GAN → the specific additions that become StyleGAN — rather than just listing differences.

Answer

Step 0: traditional GAN

Generator takes z ~ p_z, pushes it through a stack of transposed convolutions straight to an image; trained against min_G max_D V(D,G) = 𝔼_{x∼p_data}[log D(x)] + 𝔼_{z∼p_z}[log(1 − D(G(z)))]. Single injection point, entangled z, no scale-specific handles. Training is also notoriously unstable at high resolution — gradients from a single pass across many upsampling layers make convergence fragile.

Step 1: grow the network progressively

Before style-based generation, generators that aimed at high resolution learned to grow the network progressively — starting at a small spatial resolution (e.g. 4×4) and adding layers that double resolution once earlier layers stabilize, rather than training the full deep stack from scratch at full resolution immediately. This is the backbone StyleGAN's synthesis network builds on, and it's what first made stable high-resolution GAN training practical.

Step 2: add a mapping network — disentangle the latent

StyleGAN inserts f: Z → W, 8 fully-connected layers mapping z to an intermediate latent w. z is stuck matching a fixed Gaussian prior, which doesn't match the real shape of attribute variation in data, so Z is entangled. w has no such constraint — the network learns to warp the space so individual directions correspond more cleanly to individual attributes.

Step 3: style injection — replace single input with repeated modulation

Rather than feeding w into the network once, the synthesis network starts from a learned constant 4×4×512 tensor, and w is converted via a learned affine transform A at each resolution level into a style (y_s, y_b) that modulates normalized activations: AdaIN(x_i, y) = y_{s,i} · (x_i − μ(x_i)) / σ(x_i) + y_{b,i}. This is the step that turns "one latent, one shot" into "one latent, re-injected at every scale."

Step 4: stochastic variation — separate noise from style

Explicit per-pixel Gaussian noise is added independently at every layer, giving a dedicated channel for purely random micro-detail (hair strand placement, skin pore texture) that doesn't have to be encoded in w at all. This decouples "what identity/attributes to generate" (style) from "exactly how the random fine texture looks" (noise).

Step 5: multi-scale control falls out for free

Because style injection happens independently at every resolution, and resolution increases progressively through the network, control over visual features naturally splits by scale: coarse layers (4×4–8×8) → pose/face shape, middle layers (16×16–32×32) → finer structural features, fine layers (64×64–1024×1024) → colour/micro-texture.

Net effect on quality and diversity

• Quality: progressive growth stabilizes training at each resolution; normalization-based style injection keeps activation statistics well-behaved; the separate noise channel frees the style pathway from having to also encode random detail — all three compound into sharper, more coherent images than a traditional single-shot generator produces.
• Diversity: noise gives sample-to-sample stochastic variety even from the same w; style mixing (training with two different w's split at a random layer crossover) and the truncation trick (moving w towards the mean) give explicit, controllable diversity/quality trade-offs that a traditional GAN simply has no mechanism for.

Exam tip

If the question says "trace," examiners want the order of additions (progressive growth → mapping network → style injection → noise), not just a flat feature list — marks are often allocated per transition step.