- Researchers have developed a new AI video generator
- Generator creates several low-resolution stages before generating a high-resolution final
- AI generator has been made fully open source for individual and commercial use
A team of researchers from Peking University, Kuaishou Technology, and Beijing University of Posts and Telecommunications have made major strides in the field of AI video generation.
They have developed a new AI model, Pyramid Flow, capable of generating high-resolution virtual video imagery at 768p.
Unlike many proprietary AI models developed in recent years, the creators of Pyramid Flow have made their model open source, giving developers and users free access to the underlying technology.
A breakthrough in AI video generation?
The Pyramid Flow model takes a different approach from many of its predecessors by generating videos in multiple low-resolution stages before producing a high-resolution final output.
This method reduces the computing power needed to run the model, making it more cost-effective for users. The team claims that the model can produce a five-second video clip at 384p resolution in just 56 seconds.
Pyramid Flow is also available as open source software with the developers making the model freely accessible under the MIT License, which means anyone can download, modify, and use the code for personal or commercial purposes.
Along with the model itself, the researchers have posted several sample videos demonstrating the quality of Pyramid Flow’s output. The results show highly realistic, detailed imagery, including examples such as underwater explosions creating bubbles and splashing water. Furthermore, the datasets used to train Pyramid Flow, which consist of about 10 million short videos, are also available to the public.
The use of open-source datasets in AI video generation has been a contentious issue, with some critics claiming that such practices violate the rights of copyright holders.
The research team behind Pyramid Flow has not specifically addressed these concerns, but they do suggest that the model could be a useful tool for fine-tuning open-source material, potentially reducing the need to rely on third-party sources.
Via TechXplore
You might also like
https://cdn.mos.cms.futurecdn.net/R9cNexuHVtQBBawWY8vEGh-1200-80.jpg
Source link
