Lumiere AI by Google

Lumiere is a video generation and editing research model developed at Google that uses a space-time U-Net diffusion architecture to generate and modify video content by processing the full temporal structure of a video jointly, rather than generating frames in sequence or combining independently generated keyframes. The research contribution demonstrates several generation and editing capabilities including text-to-video generation, image-to-video animation, stylized generation, video inpainting for object removal and region modification, and cinemagraph creation where specific regions of a static image are animated while others remain still. The project's research page presents examples of each capability and the paper describes the architectural approach and evaluation comparisons with prior work. The code, model weights, and technical details are published as part of the academic research release. Compare alternatives. AI researchers and computer vision academics studying the development of video generation architectures use Lumiere as a reference point for evaluating the space-time diffusion approach against alternative methods such as cascaded frame generation. Machine learning practitioners and graduate students working on video synthesis reference the published paper and benchmark comparisons in their own research on temporal consistency, motion realism, and editing controllability. Research teams exploring video inpainting and generative editing study the cinemagraph and object removal demonstrations as examples of how diffusion-based editing can be applied across the temporal dimension. Academics reviewing state-of-the-art video generation approaches include Lumiere in comparative surveys of foundation model architectures Discover more tools.