Denoising Diffusion Probabilistic Model

The 2020 diffusion model that made the approach work well: add noise step by step, then learn to remove it step by step.

The denoising diffusion probabilistic model was proposed by Jonathan Ho, Ajay Jain, and Pieter Abbeel in 2020, and is the base version behind today's diffusion models. It defines two processes: a forward process that adds Gaussian noise to a real image a little at a time, following a fixed noise schedule, a timetable of how much noise to add at each step, until it becomes pure noise; and a reverse process, where a neural network is trained to predict, and subtract, the noise added at each step. The training objective is simple, just the mean squared error between predicted and true noise. To generate, start from pure noise and denoise step by step into a new sample. The paper used T=1,000 steps and a U-Net, reaching an FID of 3.17 on CIFAR-10. The downside is that sampling needs many steps, which later methods such as DDIM greatly cut down. Robotics' Diffusion Policy applies this same method to action sequences.

ExampleDiffusion Policy trains by adding noise to demonstrated action sequences over a 100-step diffusion process, teaching the network to gradually turn noise back into actions; at deployment it switches to DDIM and samples only 10 steps, taking about 0.1 seconds on an RTX 3080.

Also called

DDPM

Related

Diffusion Model · Denoising Diffusion Implicit Model · Noise Schedule · Denoising Steps · U-Net · Diffusion Policy

Sources

Denoising Diffusion Probabilistic Models (arXiv 2006.11239)
Diffusion Policy: Visuomotor Policy Learning via Action Diffusion (arXiv 2303.04137)