Startup company Deep Render, which is backed by the European Innovation Council and Intel, says that they are currently achieving 5x smaller file sizes, with potential for much more.
Video streaming is one of the heaviest sources of internet traffic so the potential gains from optimized video compression are vast, and in the future AI technology may supplant HEVC, AV1, VVC and other video codecs of today and tomorrow.
Deep Render says that it can help "save the internet" by using AI technology to make video compression more efficient while keeping video quality intact by "exploiting redundancies in a far more fluid and far more fine-grained" way.
The company has developed a software-based solution so that the algorithm can be updated in devices after implementation.
While that does not sound impossible, it will most likely require the hardware to feature a relatively powerful NPU (Neural Processing Unit) in addition to the CPU and GPU.
5x smaller file sizes
Deep Render claims to currently offer "5x smaller file sizes" and asserts that their codec "is already 80% better than state-of-the-art of traditional compression" without specifying which codec they are comparing it to.
They believe that in the future they can improve their AI-based video compression to be up to 50x more efficient than it is today.
Deep Render's illustration of how AI can help exploit redundancies in video compression. Photo: Deep Render
Backed by Intel and EIC
Those are some pretty wild claims but the 30-member team has already raised $9 million in funding this year. This includes $2.7 million from the European Innovation Council (EIC).
Additionally, the company has been selected for participation in Intel's Ignite startup accelerator program, which has led to the video introduction embedded below.
Deep Render is not the only company exploring AI-based video compression. The team behind the MPEG standards is currently developing an AI-based codec with the aim of releasing it before the end of this decade.
- Source: Deep Render via Tom's Hardware
