- original StyleGAN was special as it maps input code z to an intermediate latent code w, which applied to AdaIN layers
- stochastic variation helps the intermediate latent space W to be less entangled
- this paper investigates and fixes:
- a. a droplet artifact in original StyleGAN paper via a redesigned norm in generator
- b. artifacts caused by progressive growing design via a mixture of skip-connection and residual nets
- FID and P&R are useful but both based on classifier nets that shown to focus on textures rather than shapes
- PPL metric correlates with consistency and stability of shapes, this is expensive so texecuted less frequently
- a inversed projection, from image to latent space, helps to identify genrated images
- the droplet shape artifacts is suspected to be caused by AdaIN which normalizes the mean and variance of each features map separately, potentially destroying information relative to each other
- 2.2 Instance normalization revisited
- simply remove normalization makes the style cumulative rather than scale-spacific
- alternative: base normalization on the expected statistics of incoming feature maps, but without explictly forcing
- insdie each style block: changing from scaling the incoming feature (AdaIN), to scaling the conv weights, respectively
- observation is that after modulation and convolution, the output activation is scaled by the L2 norm of corresponding weights. The following normalization could thus be baked into the scaled weights
- thus the normalization and style block is all baked into one conv opration with scaled weights
- progressive growing have strong location preference for details
- excessively high frequencies in the intermediate layers, comprosie shift invariance
- drop progressive growing, modify MSG-GAN and LAPGAN
- MSG-GAN generator ouput a mipmap instead of an image
- up-sampling (bilinear filtering) and summing the contributions of RGB outputs of each resolutions