The BigGAN model is arguably the current state-of-the-art model in ImageNet generation (at the time of writing). Modifications incorporated into the model focus on the following:
- Scalability: The two architectural changes to improve scalability were introduced in order to improve the performance of the GAN, while at the same time improving conditioning by applying orthogonal regularization to the generator.
- Robustness: The orthogonal regularization that is applied to the generator makes the model responsive to the truncation trick, so that fine control of the fidelity and variety tradeoffs is possible by truncating the latent space.
- Stability: Devised solutions in order to minimize the instabilities.
Samples of photos generated by the BigGAN model at a 512 x 512 resolution are as follows:
The source for this image can be found at: https://arxiv.org/pdf/1809.11096.pdf...
