EgoPlay: Event-Triggered Video Editing for Egocentric Streams
EgoPlay edits first-person video only after a prompted event happens.
The paper describes a single video diffusion model that recognizes the trigger, leaves earlier frames alone, and applies the requested edit to the continuation. Its training set includes 106K event-triggered clip-prompt pairs from Ego4D-heavy data, including negative and multi-event cases. The authors report gains over EgoEdit and a VLM-guided detector-editor baseline on editing quality, visual quality, and background consistency, while using less than half the GPU memory. Accepted to SIGGRAPH Asia 2026 as a conference paper. ArXiv · AI/CL/LG's note
The paper describes a single video diffusion model that recognizes the trigger, leaves earlier frames alone, and applies the requested edit to the continuation. Its training set includes 106K event-triggered clip-prompt pairs from Ego4D-heavy data, including negative and multi-event cases. The authors report gains over EgoEdit and a VLM-guided detector-editor baseline on editing quality, visual quality, and background consistency, while using less than half the GPU memory. Accepted to SIGGRAPH Asia 2026 as a conference paper. ArXiv · AI/CL/LG's note
score 5