Generative AI has made significant strides in video editing, but most existing methods are limited to single-shot or short clips. Editing long videos with multiple instructions remains a formidable challenge, as naive chunking strategies often lead to entity fragmentation, severe editing hallucinations, and disrupted temporal continuity.
To address this, researchers from Hugging Face introduce the Multi-Instruction Multi-Shot Long-Video Editing (MMLVE) task, structured around three core objectives: Cross-Shot Editing Consistency (CSEC), Multi-Instruction Decoupling (MID), and Zero-Destruction on Spatiotemporal Structure (ZDSS).
Their proposed agentic editing framework leverages the synergy of Large Language Models (LLMs) and Vision-Language Models (VLMs) to achieve shot-level video decoupling and precise instruction parsing. This approach enables consistent, multi-instruction editing while preserving the original video's structure.
To evaluate the task, the team constructed MMLVE-Bench, a dataset featuring complex real-world spatiotemporal dynamics, high-density heterogeneous instructions, and sparse, random entity distributions. They also developed three MMLVE-focused evaluation metrics to assess editing quality.
Extensive experiments show that their MMLVE-Agent outperforms existing closed-source state-of-the-art approaches, such as Seedance 2.0, successfully eliminating editing hallucinations, preserving cross-shot editing consistency, and attaining seamless spatiotemporal transitions.