Successfully reported this slideshow.
Your SlideShare is downloading. ×

Progressive gan

More Related Content

Related Books

Free with a 30 day trial from Scribd

See all

Progressive gan

  1. 1. Progressive Growing of GANS for Improved Quality, Stability, and Variation Tero Karras, Timo Aila, Samuli Laine, Jaakko Lehtinen ICLR 2018 Oral
  2. 2. Method Generate Image Disadvantages Autoregressive (Pixel RNN/CNN) Sharp 1. Slow to evaluate (generate pixel by pixel) 2. Do not have a latent representation (similar to Deconv) VAEs Blurry Information bottleneck GANs Sharp 1. Only work on small resolutions 2. Hard to train Motivation
  3. 3. • Higher resolution is easier to tell difference. • Smaller batch size. • So Grow G and D progressively Why it is hard?
  4. 4. weight α increases linearly from 0 to 1.
  5. 5. Minbatch standard deviation • Insert a constant feature map • Compute std and concat.
  6. 6. Equalized Learning Rate • Using (He et al., 2015) initialization
  7. 7. Pixelwise feature vector normalization in G • After every conv layer (In G) • It is like batch-norm but works on pixel wise • N is the number of feature maps • In instance-norm a-mean(a) /std(a)
  8. 8. Assessment • MS-SSIM (Gobal image) (all generated images) • Proposed method (Local image structure) (generated & real) • Feature (like Sift). • Distance (using Wassertein distance Rabin et al. 2011)
  9. 9. Details • Basic loss: Improved WGAN • Model: See Right
  10. 10. Related work: Laplacian GAN
  11. 11. Generative Multi-Adversarial Networks

×