Karras, Laine, Aila, NVIDIA (2018) — ARXIV-1812.04948
Key takeaways
- Table 1 (FFHQ FID, lowest over training): Progressive GAN 8.04 → tuning, including bilinear resampling, 5.25 → mapping network and styles 4.85 → constant input 4.88 → noise inputs 4.42 → mixing regularization 4.40. The largest single step is the improved baseline, not the style architecture
- The architecture: an MLP maps
ztow. Learned affine maps ofwgive per-layer styles applied by AdaIN. The synthesis network starts from a learned constant. Per-pixel noise is added at each layer - It is still progressively grown, inheriting Progressive GAN’s networks and hyperparameters
- Also introduced the FFHQ dataset, and perceptual path length (PPL) and linear separability as measures of latent-space quality
Standing in the anthology
Filed from #163 (StyleGAN/2/3) as the base of the line StyleGAN →
StyleGAN2 (LIT-560) → StyleGAN3 (LIT-559). The practices are
filed from the later two. What they fix here is AdaIN’s droplet artifacts
and progressive growing’s phase artifacts.
Measured against diffusion. Dhariwal and Nichol (LIT-699) use StyleGAN
as the LSUN Bedroom baseline at 256², where their diffusion model scores FID
1.90 against StyleGAN’s 2.35, with precision 0.66 against 0.59 and recall
0.51 against 0.48. Their architecture ablation also tests the 1/√2
residual scaling they attribute to both StyleGAN papers, and finds it makes
FID slightly worse.
Skimmed: abstract, introduction and Table 1. No NOTE.