Text-prompt video editing has become easier with AI, but most zero-shot methods are designed for global or single-object changes. When multiple objects must be edited only in their localized regions, existing techniques often produce unfaithful results or editing leakage, where changes spill into unintended areas. PRIMEdit is a zero-shot framework that targets these problems directly.

The method introduces two key modules: Instance-centric Probability Redistribution (IPR) for precise localization and faithful editing, and Disentangled Multi-instance Sampling (DMS) to prevent editing leakage. Together, they aim to keep each edit confined to its target object while preserving the rest of the video.

To support this task, the authors also contribute the MIVE dataset for multi-instance video editing, with subsets organized by instance counts and sizes for controlled analysis. They propose the Cross-Instance Accuracy (CIA) Score as a new metric for evaluating editing leakage, reporting strong correlation with human judgment. In qualitative, quantitative, and user studies, PRIMEdit is said to significantly outperform recent state-of-the-art methods in faithfulness, accuracy, and leakage prevention. The work appears in IEEE TPAMI. Note: this digest is based on a single source, so the claims reflect the authors' own reporting.} {