StyleGAN's intermediate W space scores as more linearly separable (lower separability score) than its input Z space

- In StyleGAN, a learned 8-layer MLP maps the input latent space Z to an intermediate latent space W; the paper fits a linear SVM per binary attribute (40 CelebA-HQ attributes, labels from auxiliary classifiers on 100,000 generated images) and reports a lower separability score (conditional-entropy based; lower is better) in W than in Z (FFHQ: 3.61-3.79 for style-based W vs. 10.78 for the traditional generator's Z). [karras-etal-2018-stylegan] - Perceptual path length is shorter in W than in Z for the style-based generator with noise inputs (full path 200.5-234.0 vs. 412.0 for the traditional generator's Z), although the configuration without noise inputs has a longer full-path length in W (446.2) than in Z (412.0). [karras-etal-2018-stylegan] - Varying the depth of the mapping network (0, 1, 2, 8 layers) changes FID, path length and separability, with deeper mapping networks generally performing better; this is an architectural comparison on trained FFHQ generators (1024x1024), not an activation-level causal intervention. [karras-etal-2018-stylegan]