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.

Neural Networks · Landmark Neural Network A computational model inspired by biological neural networks, consisting of interconnected nodes (neurons) organized in layers that process information through weighted connections. Foundations · Reference Latent Variable Hidden or unobserved variables in a model that influence observed data but aren't directly measured. Vision & Multimodal · Landmark Diffusion Model A generative model that learns to denoise data, achieving state-of-the-art image generation (Stable Diffusion, DALL-E 2).

Starting from scratch? The full route 4 ideas · basics first

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.
Neural Network · read first ✓ understood
A computational model inspired by biological neural networks, consisting of interconnected nodes (neurons) organized in layers that process information through weighted connections.
Latent Variable · read first ✓ understood
Hidden or unobserved variables in a model that influence observed data but aren't directly measured.
Diffusion Model · read first ✓ understood
A generative model that learns to denoise data, achieving state-of-the-art image generation (Stable Diffusion, DALL-E 2).

Nearby papers

Canon · 1958 The Perceptron: A Probabilistic Model for Information Storage and Organization in the Brain It is the first learning neural network: the idea that weights, not rules, should hold the knowledge starts here. Canon · 1986 Learning Representations by Back-Propagating Errors Backpropagation is how essentially every neural network, from LeNet to today’s LLMs, is trained. Canon · 1998 Gradient-Based Learning Applied to Document Recognition It is the blueprint for the convolutional network, the architecture that later cracked computer vision. Canon · 2014 Generative Adversarial Networks GANs launched the modern era of generated images and dominated image synthesis until diffusion arrived.