Kiwi-Edit: Versatile Video Editing via Instruction and Reference Guidance

Researchers introduce Kiwi-Edit, a new video editing architecture that combines instruction-based and reference-guided editing for more precise visual control. The team created RefVIE, a large-scale dataset for training, and achieved state-of-the-art results in controllable video editing through a unified approach that addresses limitations of natural language descriptions.

• →Kiwi-Edit introduces a unified architecture that synergizes instruction-based and reference-guided video editing for enhanced precision.
• →The researchers developed RefVIE, a large-scale dataset created through a scalable data generation pipeline using image generative models.
• →The new approach addresses the limitation of natural language in describing complex visual nuances for video editing tasks.
• →RefVIE-Bench was established as a comprehensive evaluation framework for instruction-reference-following tasks.
• →The model achieves state-of-the-art performance through progressive multi-stage training curriculum with all code and datasets released publicly.