What is style-based generation in StyleGAN? Explain the mapping network, intermediate latent space, and style modulation. Discuss how coarse, middle-level, and fine image attributes can be controlled.
What's actually being asked
Define "style-based generation" and walk through its three building blocks — mapping network, W space, style modulation — then explain why that architecture naturally splits image attributes into coarse/middle/fine bands.
Answer
Style-based generation, in one line
Instead of feeding a latent vector into the generator once, StyleGAN starts the synthesis network from a learned constant input and repeatedly injects a transformed latent — called a style — at every resolution level, via normalization. The diagram above shows the full path from z to image.
The mapping network
z ~ N(0,I) (512-dim) is passed through 8 fully-connected layers: f: Z → W, producing w (also 512-dim). Simple shabdo mein, this is a learned "un-warping" step. z is forced into a fixed Gaussian shape that doesn't match how real attributes (age, pose, hairstyle) actually vary; w has no such constraint, so the network can bend the space until moving along one direction in w tends to change one semantic attribute more cleanly.
The intermediate latent space W
W is just the output space of f. It matters because it's less entangled than Z — not sampled from a fixed prior, but learned to approximate the true underlying manifold of variation in the training data. This is what makes attribute-level control and later latent-space editing (interpolation, direction-finding) work much better than doing the same thing in Z.
Style modulation
w doesn't go into the synthesis network directly either. At each resolution level i, a separate learned affine transform A_i converts w into a style (y_{s,i}, y_{b,i}), which modulates that layer's normalized feature maps via Adaptive Instance Normalization:
AdaIN(x_i, y) = y_{s,i} · (x_i − μ(x_i)) / σ(x_i) + y_{b,i}
where μ(x_i), σ(x_i) are the per-channel mean/std of that layer's activations before modulation. Separately, independent per-pixel Gaussian noise is added at each layer, controlling stochastic micro-detail without touching the semantic style.
Coarse, middle, and fine control
Because style is injected independently at every resolution, and resolution grows progressively through the synthesis network (4×4 up to 1024×1024), different resolution bands end up controlling different levels of detail:
Middle (16×16–32×32): finer structural features — eye shape/openness, nose and mouth structure.
Fine (64×64–1024×1024): micro-details — colour scheme (skin, hair, eye colour), lighting, fine texture.
This separation is reinforced during training by style mixing: two different w vectors are used for layers before vs. after a random crossover point, which regularizes the network away from assuming adjacent styles are correlated. At inference, the same trick lets you combine, say, person A's coarse style with person B's fine style.
Exam tip
Always name the formula for AdaIN and state explicitly which resolution ranges map to coarse/middle/fine — examiners frequently ask you to classify a given attribute (e.g. "eye colour") into one of the three bands.