Google has unveiled its new Lumiere AI for video creation. This is not a commercial product, but a research project to test innovative technologies. At the current stage, it has no restrictions imposed by legislation and social norms on the work of generative networks, and therefore the result is surprisingly good.
Lumiere is based on a “spatio-temporal U-net architecture” that forms the structure of the video throughout its entirety. This is fundamentally different from the classic approach where the AI generates the start and end frames and then tries to guess what to fill the space between them with. Lumiere creates videos that are only 5 seconds long, but they’re remarkably realistic – or surreal, depending on the initial settings.