If one trusted numbers like this, and followed a chain of en vogue codecs back through history, you'd expect that a modern codec would produce files sizes like 3% of MPEG-2 on the same input data. It's all spin.

I'm sure it does better, but I'm equally sure it'll turn out to be an incremental benefit in practice.

> I wonder how many more CPU instructions are required per second of decoded video compared to HEVC.

CPU cycles are cheap. The real cost is the addition of yet another ?!@$!!#@ video codec block on every consumer SoC shipped over the coming decade.

Opinionated bile: video encoding is a Solved Problem in the modern world, no matter how much the experts want it to be exciting. The low hanging fruit has been picked, and we should just pick something and move on. JPEG-2000 and WebP failed too, but at least there it was only some extra forgotten software. Continuing to bang on the video problem is wasting an absolutely obscene amount of silicon.

MPEG-2 doesn't support video larger than 1920x1152, so it's hard to compare on 4k video, let alone 8k. But according to https://www.researchgate.net/publication/321412719_Subjectiv... H.265 can achieve similar visual quality with 10% of the bit rate of MPEG-2 even on SD video (832*480). [Edit: not 720p]

832x480 is ~480p video (720x480 is standard widescreen dvd) 720p is 1280x720

Video sizes are not shrinking. 8K is coming, with even bigger sizes on the way. That matters at scale.

As long as someone sees the benefit, people will keep pursuing it. Compression (as like all tech) is a moving target with the platforms regularly improving on many axes.

The video codec battle is not about file size, it is about streaming bandwidth.

Are those different? If I chop up a video file into chunks, I'm streaming it, and if I save a stream I have a file. With a buffer, I would expect the sizes involved to be identical. (Although without a buffer I'd expect streaming to be worse)

Yes. For the most part, you need to be able to encode to a low and constant bandwidth at 30/60/+ fps, possibly even with limited latency. Then there are also some lesser, but also important aspects, such as the need to be able to start in the middle of a stream, handle lost packages etc.

Latency will become more important with video conferencing (already now) and AR/VR.

>Are those different?

No. In the end you transmitted a file that has a certain size. It doesn't really matter if you save that file or just use a volatile buffer.

>some objective measure of fidelity to the original

This is part of the problem. What is an "objective" measure of perceptual fidelity to the original?

A measure that has consistently high agreement when evaluated by humans, repeatable across laboratories. Image quality tests can (and are) be conducted in a structured fashion.

No such measure exists, at least not one that can’t be gamed.

No if you allow / don't control for attempts to game it then no such metric can be made. One attempt to counter this is to use multiple independent labs.

SSIM

I wouldn't say so, see Figure 1:

https://openaccess.thecvf.com/content_cvpr_2018/papers/Zhang...

It's not perfect but much better than nothing

Usually several test sets are created periodically to handle any type of situation in a variety of resolutions and depending on which part of the codec people are working on, this process is very complicated and take a lot of time to ensure that all the kind of content is representative. The metric is not optimal for video content, but gives an idea if they are going in the right direction. Ultimately, blind tests with people (experts and non experts) are done on the content to have subjective measures added to the objective ones

This has been going on since the birth of video codec. So for anyone who is new to this, the 50% are under the best case scenario. So they will compare it in 4K Res, where VVC has the highest advantage against HEVC, in measurement that fits those numbers. ( PSNR or SSIM ).

For lower res you will see lower percentage. You see the same claims from other codec as well.

Those numbers are valid for a specific PSNR... but marketing and PR obviously prefer to skip that point.

For comparison, HEVC claimed 50 to 60% compared to AVC. You can compare with reality...

PSNR is a terrible way to measure video quality. SSIM is better but not great. VMAF is good but it’s skewed towards the types of material netflix has.

Objective video quality is a tricky thing to do and you can easily come out with a codec that’s great at fine details in foreground people but terrible at high speed panning of water and trees. Depending on the material depends where you want the quality to go.