Source
Karras et al. (2019), LIT-560 — StyleGAN2. Read as NOTE-301.
Filed for #163’s “progressive training”.
The practice
Do not grow the networks during training. Progressive growing, adding resolution blocks as training proceeds, stabilized early high-resolution GANs. It also leaves details that stay glued to image coordinates as content moves (teeth that do not turn with the head), because every intermediate resolution was once trained as the final output.
Instead, fix the topology from the start:
- Generator with output skips: each resolution block emits an RGB image, and the outputs are upsampled bilinearly and summed. Early in training the low-resolution outputs dominate, so the coarse-to-fine emphasis happens by itself
- Residual discriminator, with the downsampled image fed at each resolution
| FFHQ FID (no growing) | D original | D input skips | D residual |
|---|---|---|---|
| G original | 4.32 | 4.18 | 3.58 |
| G output skips | 4.33 | 3.77 | 3.31 |
| G residual | 4.35 | 3.96 | 3.79 |
In the main ablation this change takes FFHQ FID from 4.34 to 3.31.
Both the diagnosis and the replacement come from StyleGAN2, LIT-560, which argued that progressive growing causes the phase artifacts and first recommended the fixed skip/residual topology in its place. The grid above is its Table 2, run with growing disabled; the 4.34 → 3.31 step is row E of its Table 1, where removing growing gave the largest single FID gain in the sequence of changes.
Conditions
- StyleGAN’s modulated generator. The grid was run on it
- Dataset-dependent winner. On LSUN Car the best pair is residual G with residual D (2.66). The skip-G/residual-D choice is FFHQ’s
- Single training runs, snapshot selected by lowest FID
- The skips are the replaceable part. StyleGAN3 (LIT-559) keeps the fixed topology but removes the output skips, on the hypothesis that their benefit was mainly gradient-magnitude dynamics, which it handles with an EMA normalization instead. What survives both papers is no growing, not the particular skip wiring