ML//diffusion model//consistency model

A consistency model is a model distilled to map any noise level directly to the clean image in one step. Standard diffusion (DDPM, whose denoising process it distills) needs 20 to 50 denoising steps; a consistency model needs 1 to 4, which makes it orders of magnitude faster.

A consistency model is a model distilled to map any noise level directly to the clean image in one step. Standard diffusion (DDPM, whose denoising process it distills) needs 20 to 50 denoising steps; a consistency model needs 1 to 4, which makes it orders of magnitude faster.

The core tradeoff is to sacrifice some quality for a large gain in speed. Latent Consistency Models (LCMs) brought real-time generation and the viral quick draw moment of December 2023, and the approach was later updated with sCMs and DMDs.