3 ms·
Shouldn't deep learning based video codecs take over dedicated hardware video decoders as more tensor cores become available in all new hardware? NVIDIA's DLSS
by xiphias2 6y ago
Shouldn't deep learning based video codecs take over dedicated hardware video decoders as more tensor cores become available in all new hardware?
NVIDIA's DLSS 2.0 supersampling is already moving into that direction.
- sp332 6y agoInstead of a video file or stream, that would be more like shipping a program that recreates the video. It might be cool, but it's not really feasible to play back that kind of thing on normal TV hardware.
- xiphias2 6y agoI'm not sure what you mean. There are already multiple research articles that show that deep neural network based video compression can be competitive, here's an example: https://papers.nips.cc/paper/9127-deep-generative-video-compression.pdf https://papers.nips.cc/paper/9127-deep-generative-video-comp...
- sitkack 6y ago> it's not really feasible How do you know this? TV hardware is on par with browsers. Anything is a program.
- somethingsome 6y agoIt's surprisingly difficult to guarantee a high quality on every kind of videos and formats using a neural network. Furthermore, the network should be able to handle all the corner cases (think about color profiles alone..)