Efficient Diffusion Models for Image Generation: A Cost-Decomposition Perspective

Abstract

Diffusion models have become a dominant approach to image generation, but their iterative denoising process still creates high inference latency, memory consumption, and deployment cost. Existing surveys summarize efficient diffusion models from the viewpoints of sampling, compression, architecture, or applications, but they often do not make explicit which part of the inference cost is reduced by each method. This short survey reorganizes efficient image diffusion methods through an inference cost-decomposition perspective. The overall inference cost is described by three terms: the number of denoising function evaluations, the cost of each denoising step, and auxiliary overhead from conditioning, decoding, memory movement, and deployment frameworks. Based on this view, representative methods are grouped into sampling-step reduction, few-step student models, per-step computation reduction, and system-level deployment optimization. The discussion highlights that efficient diffusion is not simply a matter of using fewer sampling steps. Practical acceleration requires balancing image quality, latency, memory, training cost, and hardware constraints. This paper provides a compact framework for comparing efficient image diffusion methods and for selecting acceleration strategies under different deployment scenarios.