Google Announces Own Video Generating AI: Imagen Video

California: In a bid to not be outdone by Meta’s ‘Make-A-Video’, Google, on Thursday, revealed its own AI video generating system dubbed as Imagen Video. This artificial intelligence (AI) system can create video clips given a text prompt (for example, “A happy elephant wearing a birthday hat walking under the sea.”) Google claims that Imagen Video is a step toward a system with a “high degree of controllability” and world knowledge, including the ability to generate footage in a variety of artistic styles. The results aren’t perfect; the looping clips the system generates frequently have artefacts and noise.

A few video clips generated through Imagen Video, as claimed by Google, were shared by the tech giant on the website.

Video Source: Google. “A happy elephant wearing a birthday hat walking under the sea.” as created by Imagen Video.

Google’s Imagen, a method for creating images that are analogous to OpenAI’s DALL-E 2 and Stable Diffusion, is the foundation for Imagen Video. A “diffusion” model, such as Imagen, creates new data (such as movies) by learning how to “destroy” and “recover” a large number of existing samples of data. The model grows stronger at recovering the data it had previously destroyed to produce new works as it is given the existing samples.

In order to generalise to a variety of aesthetics, Imagen Video was trained using 14 million video-text pairings, 60 million image-text pairs, and the publicly accessible LAION-400M image-text dataset, according to Google. Unsurprisingly, LAION was partially utilised to train Stable Diffusion.

In tests, they discovered that Imagen Video could produce videos that resembled watercolour and Van Gogh paintings. They assert that Imagen Video demonstrated an understanding of depth and three-dimensionality, which allowed it to produce videos like drone flythroughs that rotate and capture objects from various angles without distorting them. This is perhaps even more impressive.

That does not imply that Imagen Video has no restrictions. Even the clips selected from Imagen Video are shaky and warped in certain places, as Guzdial alluded to, with things that merge together in physically impossible ways, just like with Make-A-Video.

Google claims, Imagen Video, is capable of rendering HD videos (1280×768) @24 FPS.