Vision & Multimodal Jun 2020

Denoising Diffusion Probabilistic Models

Jonathan Ho et al. · NeurIPS 2020

arXiv:2006.11239

In short

Gradually add noise to an image until nothing is left, then train a network to undo one small step of that noise at a time. Sampling runs the chain backwards from pure noise, and the authors show it produces images rivalling the best GANs.

Why it matters

DDPM is the recipe behind Stable Diffusion, DALL·E, Midjourney and today’s video models.

Read first

The 3 Field Guide ideas this paper leans on.

Starting from scratch? The full route 4 ideas · basics first
  1. Machine Learning ✓ understood

    Building systems that learn patterns from data instead of following hand-written rules, getting better at a task as they see more examples.

  2. Neural Network · read first ✓ understood

    A model built from layers of simple units, each taking a weighted sum of its inputs and applying a nonlinearity; training adjusts the weights so the whole stack maps inputs to useful outputs.

  3. Latent Variable · read first ✓ understood

    Hidden or unobserved variables in a model that influence observed data but aren't directly measured.

  4. Diffusion Model · read first ✓ understood

    A generative model trained to remove noise a step at a time, so it can turn pure random noise into a new image, audio clip or other sample.

Nearby papers

Summary in our own words; read the paper for the details. ← All papers