Research questionHow can video-editing benchmarks measure instruction fidelity across video-specific editing tasks?Many evaluations overlook temporal, audio, and reference-based editing requirements by treating video edits like frame-level image edits. They can also give high scores to visually plausible results that fail to follow the instruction. Latest papersRecent research connected to this question, newest first.OmniEdit-Bench: A Comprehensive Benchmark for Instruction-based Video EditingThe source concerns instruction-based video editing across spatial, temporal, audio, and reference-based tasks, including explicit, implicit, and reasoning-based instructions. Its benchmark assesses accuracy, preservation, realism, and consistency using human judgments and vision-language models, with experiments on representative open-source and commercial systems.research paper · Sep 2, 2026