Hierarchical VAEs: Deep Latent Structures

Module: Generative AI | Difficulty: Advanced

NVAE Architecture

Loss:

Channel-Wise Hierarchy

Latents at different resolutions:

Key Results

| Model | FID (CelebA-HQ) | Parameters | |-------|-----------------|------------| | VAE | 45.2 | 35M | | NVAE | 8.3 | 258M | | Very Deep VAE | 3.8 | 450M |

Research Insight: Hierarchical VAEs match GAN quality while maintaining exact likelihood computation. The key is multi-scale latent variables that capture both global structure and local detail.