cd ~/articles
How It Works2025-02-1910 min read

What Are Diffusion Models? A Quick Tour

U-Net encoders and decoders, convolution versus transposed convolution, PSF and FFT, forward and reverse diffusion, checkerboard fixes, classifier-free guidance, and where the VAE fits.

AI

Diffusion models are a fascinating blend of probability, signal processing, and deep learning.

What it covers

  1. 1. U-Net
  2. 2. Convolution vs. Transposed Convolution
  3. 3. PSF & FFT
  4. 4. Probabilistic Models & De Moivre
  5. 5. Forward & Reverse Diffusion
  6. 6. Checkerboard Fixes
  7. 7. Classifier-Free Guidance
  8. 8. Variational Autoencoder (VAE)

Pulled quotes

  • Forward diffusion gradually adds noise to data, step by step. Reverse diffusion learns to remove noise at each step, reconstructing the original data.

Read the full article

This article was first published on LinkedIn. The complete text, with figures, lives there.

Read on LinkedIn