Advertisement
Need for speed: Chinese researchers unveil new technique for near-instant AI video creation
TurboDiffusion is expected to enable commercial-scale video creation at less cost and a faster pace
Reading Time:2 minutes
Why you can trust SCMP
1
Ben Jiangin Beijing
A new technique called TurboDiffusion claims to be able to accelerate artificial intelligence video-generation speeds by up to 200 times at scale, without compromising overall image quality, according to a recently published paper.
The technique reduced the time needed to generate a five-second, standard-definition video clip from more than three minutes to just 1.9 seconds – a nearly 100 times acceleration – when it was tested on a consumer-grade system with Nvidia’s RTX 5090 graphics processing unit, according to researchers from Beijing’s Tsinghua University, AI model developer Shengshu Technology and the University of California, Berkeley, in a paper published late last week.
The time it took to generate a high-definition, five-second video on the same machine was cut from nearly 80 minutes to 24 seconds, or about 200 times faster, according to one experiment cited in the paper.
Advertisement
Currently, it takes between three and five minutes for a user to generate a standard five-second video on Shengshu’s Vidu platform. OpenAI had also said its text-to-video model Sora takes several minutes to render a short clip.
TurboDiffusion underscores the rapid strides made by Chinese researchers in advancing AI video-generation technology since OpenAI introduced Sora in February 2024.
Advertisement
Accelerated video creation represents a significant shift, as the use of video-generation models would no longer be a bottleneck in the production process, according to industry analyst Kyon Xu.
Advertisement
Select Voice
Choose your listening speed
Get through articles 2x faster
1.25x
250 WPM
Slow
Average
Fast
1.25x