HKUST AI Film Festival

In this talk we examine the different ways humans and AI systems can create together. While not strictly necessary, we may wish that AI creative partners themselves possess the quality of creativity. We look at what sort of attributes we should expect in human-level creative systems. I conclude the talk with some informal analysis of recent progress toward AI systems that express creativity.

Video creators always chase the latest technological innovations. Editors and supervisors consider mastering new workflows to be essential knowledge. But when it comes to learning storytelling skills, our industry grasps for a center of gravity. Is film school worth the investment? Do all editors need a mentor? Do YouTube tutorials hold the answers?
Until recently, creative editing was an amorphous concept taught by people promising to "tell you everything they don't in film school." These instructors offer no measurable outcomes, no metrics to track progress or assess skill level.
In this presentation, EditMentor CEO Misha Tenenbaum shares what he and his team believe is the future of learning creative video editing -- interactivity, gamification, and the use of data.

Seyhan Lee has become the studio that bridges the worlds of cutting-edge ML research and the world of entertainment. This talk will showcase how we use generative models in the production pipeline, i.e., award-winning films, music videos, and immersive experiences. and the latest research from Seyhan Lee's creative lab. I will also give a glimpse of the tool we are developing to integrate generative A.I. into the Unreal Engine for virtual production.

At Prime Video, we are always striving to create best-in-class video watching experiences for our customers. This requires us to develop novel AI technologies for video understanding and synthesis that can enable a variety of delightful video watching experiences in an efficient and scalable manner. A key challenge to this end is the large amounts of labelled training data required by the end-to-end supervised approaches that have traditionally been used for video understanding. In this talk, I will provide an overview of some of the self and weakly supervised approaches we have developed that can effectively address this challenge. These approaches can learn general purpose video representations in a label efficient manner and offer state-of-the-art performance for a wide variety of video understanding tasks. Similarly, in the space of video synthesis, the special effects workflows traditionally employed in the industry are highly manual and therefore not easily scalable. I will present an overview of some of the state-of-the-art methods we have developed to automate traditional special effects workflows, and will share results of these approaches on both cinematic as well as live-sports video content.