Video Editing Meets Diffusion Models: A Comprehensive Survey

The rapid advancement of diffusion models has revolutionized video editing, yet existing surveys overlook how foundational model architectures fundamentally shape editing paradigms. This survey provides comprehensive analysis through a novel foundational-model-centric perspective, revealing a striking pattern: 100\% of T2I-based methods prioritize temporal consistency to compensate for absent motion priors, while 100\% of T2V-based methods focus on conditional injection precision leveraging native temporal understanding, with hybrid approaches showing 68\% emphasis on conditioning. We organize our analysis around three pillars. First, we trace the technical evolution of multimodal conditional injection, progressing from text-only conditioning through hybrid multi-modal integration to unified frameworks, alongside temporal consistency mechanisms evolving from sparse attention extensions through explicit feature correspondence to comprehensive learned modeling. Second, we examine how architectural choices fundamentally shape research priorities and method design. Third, we taxonomize approaches across multimodal conditions, editing tasks, and training paradigms. We collect datasets, evaluation metrics, and benchmarks, then identify ten critical open challenges. This survey illuminates the deep connection between foundational architectures and research directions, providing a roadmap for advancing video editing capabilities.