I mean, that would work. It's going to be the reason AI video compression works; people are okay with an AI model being 1GB but they wouldn't put up with libavcodec being 1GB.
It's a valid codec move to have 1GB in the codec but to be able to compress arbitrary video with it, or even just arbitrary video within a certain specialized domain. Having those requirements will affect all the cost/benefit decisions that get made when people decide whether to use it, but if it outperforms on other metrics it may be something that wins in some places.
I believe lumost is referring to the actual video being used for testing being embedded in the codec. That is not a valid move; it compresses just that one exact video arbitrarily small (honestly anything above zero bytes is just sandbagging, you can always map the empty file to your test video, for INFINITE COMPRESSIONS!!!1!) but nothing else.
>> people are okay with an AI model being 1GB but they wouldn't put up with libavcodec being 1GB.
If people are happy with the results of the libavcodec, you could rebrand it as "libavcodec-ai" and now you have a more effective codec that might be bigger, but is now palatable to users :-)
