Recent advances in video generation have outpaced progress in video editing, which remains constrained by several limiting factors, namely: (a) the task's dependency on supervision severely limits generality, (b) an unnecessary artificial separation between the generation and editing task, and (c) the high computational costs of training a video model. In this work, we propose UES (Unlocking Universal Editing via Self-Supervision), a lightweight self-supervised fine-tuning strategy that transforms generation models into unified generation-editing systems through self-supervised semantic alignment. Our approach establishes a dual-conditioning mechanism where original video-text pairs jointly provide visual and textual semantics, enabling structured learning of intrinsic spatiotemporal correspondences. Key advantages include: (i) Universality through supervision-free adaptation to diverse editing tasks, (ii) Unification of generation and editing applicable to most text(+image)-to-video model, and (iii) Efficiency via lightweight fine-tune that reduces tunable parameters by 92.67%. To enable systematic evaluation, we introduce OmniBench-99, a comprehensive benchmark spanning 99 videos across humans/animals, environments, and objects, comprising 4 editing types and 8 scenarios. Extensive experiments show UES enables models without inherent editing capability to perform powerful and universal editing while preserving or even enhancing their original generation performance.
翻译:暂无翻译