How the New OLSS Algorithm Supercharges Diffusion Model Sampling

Paper Selected at CIKM 2023

Recently at CIKM 2023, a paper titled “Optimal Linear Subspace Search: Learning to Construct Fast and High‑Quality Schedulers for Diffusion Models” authored by Duan Zhongjie, Wang Chengyu, Chen Cen, Huang Jun, and Qian Weining, and developed by Alibaba Cloud AI platform PAI together with East China Normal University, was selected.

The proposed OLSS (Optimal Linear Subspace Search) algorithm treats diffusion‑model sampling acceleration as an expansion process in a linear subspace, offering a unified analysis of existing methods and introducing a new scheduler that significantly speeds up image generation.

Why Diffusion Model Acceleration Matters

Diffusion models have become a dominant technique in image generation, powering systems such as Stable Diffusion, Midjourney, and many open‑source projects. However, their iterative denoising process requires many model evaluations, leading to high computational cost.

Existing acceleration approaches design “schedulers” that approximate the full generation trajectory with fewer steps. The paper shows that these approximations are essentially linear‑subspace expansions; DDIM operates in a two‑dimensional subspace, while OLSS searches in higher‑dimensional subspaces for better approximations.

OLSS also incorporates a path‑planning component that further improves fidelity, achieving higher image quality at the same number of steps compared with prior schedulers.

Paper Details

Title: Optimal Linear Subspace Search: Learning to Construct Fast and High‑Quality Schedulers for Diffusion Models

Authors: Duan Zhongjie, Wang Chengyu, Chen Cen, Huang Jun, Qian Weining

PDF: https://arxiv.org/abs/2305.14677