Thinking on Shots: Consistent Multi-Shot Video Editing with Agentic Reasoning
The paper proposes an agent-based framework for editing long, multi-shot videos without losing continuity.
The authors define a new task, MMLVE, aimed at handling multiple edit instructions across long videos. Their approach uses LLMs and VLMs to split video at the shot level and parse instructions more precisely. They also introduce MMLVE-Bench and task-specific metrics for consistency, instruction separation, and structure preservation. In experiments, they report better results than closed-source systems including Seedance 2.0. HF Daily Papers' note
The authors define a new task, MMLVE, aimed at handling multiple edit instructions across long videos. Their approach uses LLMs and VLMs to split video at the shot level and parse instructions more precisely. They also introduce MMLVE-Bench and task-specific metrics for consistency, instruction separation, and structure preservation. In experiments, they report better results than closed-source systems including Seedance 2.0. HF Daily Papers' note
score 4