A narrower code forces stronger compression. With no nonlinearity, the autoencoder reduces to PCA.
Computed: 60,000 CelebA faces at 32 × 32, grayscale; R2 on 2,000 held-out faces.