1. Introduction

• 2D GAN
  • lack 3D understanding
  • multi-view consistency X
• NeRF variants
  • GRAF, pi-GAN
  • GIRAFFE
  • 3D consistency X , high-resolution X

2. Related Work

• StyleGAN
• StyleGAN2

3. Method

3.1 Image Synthesis as Neural Implicit Field Rendering

Challenges

• computationally expensive
• Difficult to train on high-resolution results

3.2 Approximation for High Resolution Image Generation

• 2D GANs generate images fast because
  • Each pixel only takes single forward pass through the network
  • Image features are generated progressively (coarse → fine)
    • High resolution feature maps : small number of channels
• StyleNeRF
  • Aggregating features into 2D space before final colors are computed

3.3 Preserving 3D Consistency