Social Media Video Creation: Harnessing the Power of Large Language Models for Non-Linear Editing and RAG-Based Question-Answering

This paper presents a multimodal tool, Video Swagger, the first framework for real-world voice backed video editing for businesses preparing marketing videos to be posted on social media platforms with non-linear output. This is a comprehensive tool that allows automation of the video editing and after the video is uploaded on the social platforms, and further, takes care of generating the automatic responses for the users' questions using RAG, that further enhances the tool. While LLMs have proven effective for various tasks, this model harnesses their power to create short and effective video posts for social media platforms. There are currently no video editing models publicly available that can arrange the various snippets in a long video in such an order that it conveys the audience effectively in less than 60 sec. We evaluated the performance of Video Swagger using human raters: the performance of the model was assessed by expert human video editors on a seven-point Likert scale using an eight-point metrics. The evaluation showed strong correspondence to the standards of the experts, proving that the model can create relevant and engaging videos within a fraction of the time compared to human-based editing. The results point out the model as a potential comprehensive tool for video editing.