English

News

Mechanisms and Challenges of Distillation Techniques in Reducing Diffusion Model Sampling Steps

This article is a translation. Read the Japanese original

Diffusion models handle the difficult task of generating data from high-dimensional distributions by breaking it down into many denoising tasks. The key to generation lies in making predictions at each step and repeating them continuously.

However, recent research has shown a movement toward reducing the number of sampling steps, and even aiming for single-step sampling. This appears to contradict the nature of these models, where many subdivided steps support their performance.

While addressing this paradox, the article provides a detailed explanation focusing on distillation. Distillation refers to a method of training a new model (student) using the predictions of an existing model (teacher).

The article also delves into why diffusion models require many steps to achieve high-quality results. Each step in sampling is a process of calculating a local update direction within the input space. If the step size of the update is too large, it causes the generated images to become blurry. This occurs because high-frequency information is obscured by noise, causing multiple possibilities to blend together.

--- Source: The paradox of diffusion distillation (2024) (Hacker News Frontpage, 2026-09-04)