Factorized video generation
With Emu Video, we demonstrate the implementation of factorized video generation through a single diffusion model. This breakthrough paves the way for versatile applications across diverse video generation tasks. Our research explores crucial design decisions, such as fine-tuning noise schedules tailored for video diffusion. Additionally, we employ multi-stage training strategies to empower the direct generation of higher-resolution videos, pushing the boundaries of visual fidelity.