VideoPoet is a Google research project that uses large language models to generate and edit video and audio from text, images, and existing footage.

VideoPoet is a Google research project that turns a large language model into a video generator. It is built on pre-trained MAGVIT V2 video and SoundStream audio tokenizers, which convert images, videos, and audio clips into discrete codes a language model can process. An autoregressive language model then predicts the next piece of video or audio in a sequence to generate new content.
Because it is built for multimodal learning, VideoPoet handles text, images, and video together. That enables text-to-video generation, image generation from text, and editing of existing footage while keeping it consistent over time. It produces a wide range of motion, supports multiple video formats, and can generate audio matched to the visuals.
VideoPoet was developed by a research team at Google, including Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, Yong Cheng, Ming-Chang Chiu, Josh Dillon, Irfan Essa, Agrim Gupta, Meera Hahn, and Anja Hauth.
VideoPoet suits content creators, digital marketers, video editors, social media managers, animators, graphic designers, film producers, advertising specialists, educators, and videographers — anyone who needs generated or edited video content.
VideoPoet is available through Google’s research site, which showcases its demo outputs and capabilities. Example prompts demonstrate how a single text prompt, such as an astronaut riding a horse in a forest, can be rendered in different styles, including photorealistic, digital art, and animated oil painting.
Related
Discover more tools in similar categories
Weekly picks of the best AI tools we've researched and added — no spam, unsubscribe anytime.
Join 25,000+ AI enthusiasts. No spam, unsubscribe anytime.