dorarad
commented
3 years ago Hi there, yes there're gaps between numbers reported in StyleGAN papers and our work that stem from the length of training. As noted in the StyleGAN paper bottom of page 10 they have trained their model for very long times: 25k image-steps for FFHQ, and 70k steps for Bedrooms. This is equivalent to 91 GPU-days of training on NVIDIA V100.
Therefore, we compare all approaches given the same and lower more feasible number of training steps: we explored training on different datasets in the ranges of 5k to 15k image-steps (are equivalent to about 3-4 days on 4 GPUs), about 5-7 shorter training than the experiments in the StyleGAN paper. Under these settings, our model obtains better FID scores than StyleGAN as detailed in table 1 page 8 https://arxiv.org/pdf/2103.01209.pdf. See also footnote about that aspect in the bottom of page 8.
Furthermore, to give additional insight into the difference between models' performance over the course of training, you can see on page 9 figure 7 of our paper the learning curve in terms of FID score over steps, showing how given the same number of training steps the GANsformer performs better than StyleGAN both for the Bedroom dataset and especially for e.g. the CLEVR dataset. (Note also that the x-axis there is measured in training steps rather than image-steps but I'll fix it to match the more common image-step notation)
That's an important point so I hope my explanation here helps clarifying that!
betterze
The number was reported in Figure 10 of its paper: https://arxiv.org/abs/1812.04948