AI 'assisted' video compression

The idea sprung to me awhile ago.
VEAI has the ability to use CRF to select image quality output of a video. Though I was wondering why it couldn’t just use AI to determine a better output?

The idea being the AI is only looking for artifacts between two images. These two images would come from the the AI taking multiple screenshots throughout a given video and running them through various CRF values.

It will then output a recommended CRF value based on what it thinks is most ideal for compression to maximum image quality with little to no produced artifacts.

Interesting idea, however here are my thoughts. Any value higher than CRF 0 is a compromise of quality vs. file size, so you want to give AI the task to decide at which CRF value the quality is ‘good enough’. Some people think CRF 20 looks reasonable. Personally, I never go higher than CRF 14 because I can see a degradation that others seemingly do not perceive or care about. How can AI make such a subjective quality decision? The simple answer is, it cannot. If it could, then every VEAI upscale result would be ‘ideal’, and it is certainly not the case. That is why we have all the different AI model choices that are continually being trained and improved. In the end, I would hope that we as observers and end-users should always make the final call of what is acceptable quality.

1 Like