GGDR
Created on November 16|Last edited on November 16
Comment
Quality estimation
- Base model -- Discriminator consists only of decoder, no GGDR.
- BigLambda -- Discriminator with encoder and decoder, regularization coefficient = 50
- The rest of models have regularization coefficient = 10.
Here only BigLambda and TooManyEpochs have results after step 781 (corresponds to 100k images). The rest of models show results similar to TooManyEpochs: discriminator loss becomes too small, which causes generator to fall into small model and generate only handful of different faces. Note, that BigLambda (regularization coefficient = 50) avoids this problem thanks to large enough regularization.
Run set
6
Run set
7
Performance estimation
For comparison: StyleGan2 requires 450 * 10^10 FLOPs for batch of 100 images 256 x 256. Source: https://arxiv.org/abs/2104.02244
Add a comment