Google & HUJI Present Dreamix: The First Diffusion Model for General Video Editing

Diffusion models like Stable Diffusion — which introduce random noise to data and then learn to generate new samples from the noise — have achieved state-of-the-art performance in generating realistic text-driven images and videos. Such models however are focused on synthesizing, not editing. While a number of intuitive text-based approaches have…