OpenAI's Sora could be facing its biggest competitor yet. Luma's artificial intelligence video generator, Dream Machine, has caused multiple-hour wait times after high demand for the software's free public beta version's debut.
As confirmed by the company, the startup backed by famed Silicon Valley venture firm Andreessen Horowitz says the sheer volume of traffic has caused longer video processing times while stating that they will continue to increase the software capacity.
As of January 2024, reports claim that the little-known tech startup had received over $70 million, notably $43 million from its Series B. The startup debuted its text-to-3D asset-generating model Genie 1.0 in November 2023. This is on top of the startup's excellent AI video generator.
Prominent AI video creators and filmmakers were allowed to test Dream Machine's ability to create videos from text prompts and still images before the public beta opened. Early posts proved to be positive, with others who are just getting their hands on it also finding it extremely impressive. This invites comparisons to OpenAI's Sora, while some say it is already superior.
Tests from other sources show that the text-to-video feature only occasionally depicted what was asked in a prompt. The film was produced in just a few minutes and featured high-resolution, highly detailed materials and incredibly smooth, non-jittery action.
China's Short Video App Joins AI Video Generating Race
With an app called Kling, China's second-biggest short video application, Kuaishou, most recently entered the race to create the best AI video generator. The application is said to produce high-quality videos using text cues.
Kling is currently in experimental mode. It can convert text inputs into 1080p video clips up to two minutes long. The manufacturer claims that Kling can produce realistic and fanciful scenes and supports a variety of aspect ratios. The demonstration movies showed various scenarios, such as a boy eating a hamburger and a white cat driving through city streets.
Google's Lumiere
This year, Lumiere was one of the earliest AI video generation programs published by Google. A novel Space-Time U-Net architecture is presented in this Google Research project, which can produce an entire video's temporal duration in a single model pass.
Unlike other video models that generate remote keyframes followed by temporal super-resolution, Lumiere uses a distinct technique that makes global temporal consistency more achievable.
The architecture combines temporal and spatial down- and up-sampling with a text-to-image diffusion model that has already been trained. Lumiere can now handle the data at various space-time scales and natively output a full-frame-rate, low-resolution video.
The ground-breaking space-time U-Net architecture can generate full-frame-rate video clips for uses ranging from image-to-video and video inpainting to the creation of styled content.
