Diffusion-driven key segment generation via progressive boundary optimization for partially relevant video retrieval

Partially Relevant Video Retrieval (PRVR) aims to retrieve related videos from a large collection of untrimmed videos. However, existing methods suffer from redundant candidate generation and fixed-length constraints. This paper presents the first diffusion-driven framework for PRVR, addressing two critical challenges: diffusion target ambiguity and the computation-performance trade-off. First, we redefine the diffusion target as boundary coordinates through a conditional Markov process, enabling the differentiable modeling of continuous boundary variables with geometric constraints to ensure temporal coherence. Second, we propose a progressive boundary refinement strategy that balances exploration (high noise) and exploitation (low noise) through adaptive noise scheduling. This mechanism allows for broad boundary search in early iterations and precise adjustment in later stages. Experiments demonstrate that our approach reduces redundant candidates by 99% compared to sliding-window methods and achieves 44% faster training than MS-SL on Charades-STA. The results validate the effectiveness of diffusion models in PRVR tasks.