Progressive Growing of GANs for Improved Quality, Stability, and Variation The title image is an image generated by this GAN, not a real image. In short, it’s all a computer imagination that can endlessly generate a variety of “starlike” images.
Be /XOxxPcy5Gr4, for those with access to Youtube.
By using progressively larger GAN networks (first train 4×4, then train 8×8, then… Up to 1024×1024), combined with the carefully processed Celeba-HQ dataset, to achieve the most realistic GAN effect to date.
And the network grows by using the idea of residuals:
Take a look at some more images generated by it, naturally beautiful:
To be honest, it looks a little too familiar. I suggest nVidia add the most similar image in the dataset as a comparison to determine whether it is similar to the image in the dataset.
Such a good effect, and data are inseparable. Celeba-hq dataset is generated for training by super-resolution, blurring background, and alignment:
And in the bedroom data set, it’s completely impossible to tell:
Finally, the code and network data are also exposed, see tkarras/ Progressive_growing_of_gans.
Can be in drive.google.com/drive/folde… See more hd images:
