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

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