IMAGE GANS MEET DIFFERENTIABLE RENDERING FOR INVERSE GRAPHICS AND INTERPRETABLE 3D NEURAL RENDERING

less than 1 minute read

Published:

styleGAN as a generator, extract 3D params., DR for reconstruction

  • goal: extract and disentangle 3D knowledge by generative models
  • GAN as a neural renderer
  • styleGAN: lower layers for viewpoint, while higher layers for shape, texture and background
    • fine tune with inverse graphic network fixed
    • disentangle bg and fg with seperated losses