Course contentsShow
← Machine Learning and Deep Learning
Lesson 1549 of 3,53834. Diffusion ModelsFree lesson

DDPM vs VAE: Key Differences

Comparing DDPM's iterative refinement and fixed encoder to VAE's single-step generation and learned encoder.

DDPM vs VAE: Key Differences

You'll learn: How DDPMs and VAEs take fundamentally different approaches to generating data, comparing their encoding and generation strategies.

The Core Philosophical Split

Both DDPMs and VAEs are generative models, but they solve the generation problem in opposite ways:

VAEs compress data into a learned latent space through an encoder, then reconstruct it in one shot with a decoder. The encoder learns to map each input to a compact representation, and the decoder learns to reverse that mapping.

DDPMs gradually destroy data through a fixed forward process (adding noise), then learn to reverse that destruction step-by-step. There's no encoder that "understands" the input—just a denoising network that removes noise iteratively.

Three Key Differences

1. Encoding Strategy

  • VAE: Uses a learned encoder network that compresses data into meaningful latent codes
  • DDPM: Uses a fixed forward process (noise schedule) with no learnable parameters

2. Generation Process

  • VAE: Single-step generation from latent code to output
  • DDPM: Iterative refinement through hundreds or thousands of denoising steps

3. Latent Space

  • VAE: Compact, structured latent space (e.g., 128-dimensional) designed for interpolation and manipulation
  • DDPM: Each timestep creates an intermediate noisy version—no explicit compressed representation

Think of it like art restoration: VAEs sketch a blueprint first, then paint the full picture in one go. DDPMs start with complete noise and gradually reveal the image, like slowly removing layers of dirt from an old painting.

Key Takeaway: VAEs learn both compression and generation with fast single-step sampling, while DDPMs use fixed noise corruption with learned iterative denoising for higher quality at the cost of slower generation.