VideoPoet
Google has unveiled VideoPoet, an AI-driven tool enabling creators to generate videos from text, images, and audio input. The tool boasts features like video frame continuation, inpainting and outpainting, video stylization, and video-to-audio capabilities.
It simplifies the video creation process, allowing users to generate videos, controlled motions, and stylized effects based on text prompts. Users can input prompts, and the model responds with results or can request custom motions, styles, or audio based on provided images.
Developed with the advanced MAGVIT-2, VideoPoet facilitates the production of high-motion variable-length videos. MAGVIT-2, a video tokenizer, generates concise and expressive tokens for videos and images using a shared token vocabulary. The tool also offers interactive editing features, consolidating various video generation functionalities within a single Large Language Model (LLM).
According to Google, VideoPoet is a modeling method capable of converting any autoregressive language model or large language model into a high-quality video generator.
Google clarified that VideoPoet operates as an autoregressive model, meaning it creates output by referencing its previous generation. While not yet accessible to the public, Google’s research team has released a demo website showcasing multiple demos and videos created using the platform. The tool’s versatility is demonstrated by its capability to combine various video clips to create a short film.
As of the latest update, VideoPoet is still in development and has not been made publicly available by the tech giant. However, the demo website provides a glimpse into the potential of this innovative AI-driven video generation tool.
