中圖分類號：TP391 文獻(xiàn)標(biāo)志碼：A DOI: 10.16157/j.issn.0258-7998.257410
中文引用格式： 趙家池，張杰. 擴(kuò)散模型的推理端優(yōu)化：可控性與加速的方法綜述[J]. 電子技術(shù)應(yīng)用，2026，52(6)：35-40.
英文引用格式： Zhao Jiachi，Zhang Jie. A survey of inference-time optimization in diffusion models: controllability and acceleration[J]. Application of Electronic Technique，2026，52(6)：35-40.

A survey of inference-time optimization in diffusion models： controllability and acceleration

Abstract： Diffusion models show strong performance in images, speech, and scientific computing. Practical deployment, however, faces two bottlenecks: limited controllability at inference and high inference cost. Conditional alignment can fluctuate due to sampling randomness and time-step discretization. Multi-step sampling also increases latency. This survey focuses on Inference-Time Optimization (ITO). We organize recent work along two axes: controllability and efficiency. Without changing the pretrained backbone or the data, ITO uses constraint, scheduling, and selection strategies to improve conditional alignment and detail stability under few-step or one-step budgets, while reducing latency. We also outline transferable application scenarios. Finally, we present deployment-oriented prospects that aim to improve speed, coverage, and quality in a unified manner.

Key words : diffusion models；inference-time optimization （ITO）；controllability；efficiency