CoVEBench: Can Video Editing Models Handle Complex Instructions?

Researchers introduce CoVEBench, a comprehensive benchmark for evaluating video editing AI models on complex, multi-step editing tasks. The benchmark reveals that current video editing models struggle significantly with compositional instructions that require simultaneous modifications while preserving unrelated content, exposing a critical gap between simple isolated edits and real-world user workflows.

CoVEBench addresses a fundamental limitation in how video editing models are currently evaluated and developed. While existing benchmarks focus on isolated, straightforward tasks like style transfer or single object insertion, real users demand sophisticated workflows requiring multiple coupled edits—changing subjects, actions, and camera angles simultaneously while preserving background elements. This disconnect between academic evaluation and practical application has allowed models to appear capable while actually failing at compositional reasoning.

The benchmark's scale and rigor are notable: 416 curated videos, 626 multi-point instructions, and 9,990 fine-grained checklist items enable detailed diagnostics of model behavior. By employing MLLM-judged instruction compliance alongside automated video quality metrics, CoVEBench moves beyond blunt global measurements toward nuanced assessment of what models actually accomplish versus what users request.

The experimental findings expose systemic weaknesses in current architectures. Models frequently omit edits entirely, fail to respect preservation constraints, or generate artifacts when handling multiple operations—failures that undermine practical deployment. These results validate that compositional video editing represents a genuine technical challenge requiring architectural innovations beyond incremental improvements.

For the AI development community, CoVEBench serves as both diagnostic tool and research agenda. Developers building production video editing systems now have a rigorous testbed to identify failure modes and measure progress. The benchmark's diagnostic approach should accelerate focused research on compositional reasoning in video models, ultimately advancing the field toward systems that handle realistic user demands.

• →Current video editing models fail at compositional tasks requiring multiple simultaneous edits despite succeeding at isolated operations.
• →CoVEBench's 9,990 checklist items and MLLM evaluation enable fine-grained diagnosis of model failures beyond coarse global metrics.
• →Real-world video editing demands multi-coupled operations that preserve unrelated spatiotemporal content, a capability not yet mastered by existing models.
• →The benchmark reveals models frequently omit edits, violate preservation constraints, or introduce artifacts when handling complex workflows.
• →CoVEBench establishes a rigorous research agenda for advancing video editing AI toward practical real-world applications.