Denoising Diffusion Implicit Model

A diffusion speedup method that keeps DDPM's training but turns sampling into a deterministic process that can skip steps.

Denoising Diffusion Implicit Models were proposed in 2020 by Stanford's Jiaming Song, Chenlin Meng, and Stefano Ermon, and published at ICLR 2021. A DDPM (Denoising Diffusion Probabilistic Model) generates one sample by walking a Markov chain of hundreds to thousands of denoising steps, which is slow. DDIM constructs a family of non-Markovian diffusion processes with the same training objective as DDPM, so an already-trained DDPM can switch to DDIM sampling with no retraining: it can skip steps, running only a dozen to a few dozen, which the paper reports is 10 to 50 times faster in wall-clock time; and setting the stochastic term to zero makes sampling deterministic, so the same starting noise always gives the same result. It became the starting point for the many fast diffusion samplers that followed, and Diffusion Policy commonly relies on it to cut steps for real-time control on real robots.

ExampleDiffusion Policy's real-robot experiments train with 100 steps but run inference with 10 DDIM steps, giving about 0.1 seconds of inference latency per call on an RTX 3080.

Also called

DDIM, DDIM Sampler

Related

Denoising Diffusion Probabilistic Model · Diffusion Model · Denoising Steps · Diffusion / Flow Samplers · Diffusion Policy · Noise Schedule

Sources

Denoising Diffusion Implicit Models (arXiv:2010.02502)
Diffusion Policy: Visuomotor Policy Learning via Action Diffusion (arXiv HTML)