StyleFlow Attribute conditioned Exploration of StyleGAN-Generated Images using Conditional Continuous Normalizing Flows

attribute-conditioned sampling and attributecontrolled editing

• the former as finding non-linear paths in latent space
• recall: z -> w -> image
  • for the second task, given a image and desired attributes w, we first map the image to latent space to get a w, predict its attributes a_i, invert to get the initial z_0
  • how to invert the z -> w function? CNF!
  • CNF is clamed to be less entangled
• specific editing works best at specific stages/layers
• by continuous, a time variable is introduced in each layer of CNF
  • seems in code set to 0 -> 1
  • using torchdiffeq, an ode solver