Does VMAF scores even really matter?
In OBS I encoded a source file in av1 on my arc card with an ICQ of 1 for as much quality preservation as possible. Then I went to handbrake. and re-encoded the same file in hand break with a much more sensible ICQ of 23. Putting the files into FFMetrics and running it through VMAF for about 30 seconds. VMAF is telling me that the ICQ 23 clip has a score of 75.0015. which I would assume that I should probably easily be able to see something is up with the pics without needing to pixel peep really. But I couldnt tell much of a difference at all. and only when I took screenshots and looked back and forth over and over again multiple times looking to try and find any differences. I can hardly tell that maybe the grass was just ever so slightly better on the ICQ 1 than the ICQ 23 video.
How could something that looks so close to being the best that it can be. look to what should be the best quality it can be. have such a drastically lower score? Is VMAF flawed or is how I am testing this inheriantly flawed or is it working about right?
1st and 3rd is ICQ 1 and 2nd and 4th is ICQ 23
6 Replies
and the pics sent out of order, rip
Update: I decided to throw the original ICQ source file into handbreak. at ICQ 1 and now when comparing to the ICQ 23 file it gives me a more reasonable VMAF score of 98.7177
Why is this the case?
when you're trying to compare VMAF stuff its incredibly important that they are identical in terms of timestamps and what have you.
You CANNOT mix mkv + mp4 for instance.
Obviously, you also need to make damn sure that they are lined up frame perfectly.
There are other ways where things can drift over time, and you just need to keep an eye out for it. Its usually quite apparant in the graphs/scores if something has gone awry.
You should never trust any metric completely blindly I'd say 🙂
Makes sense. So the solution was to probably remux the recording?
GG @me.when.its.spider.demon, you just advanced to level 1 !
Could be due to the remuxing or re-encoding, either one could've caused issues like Flaeri described
Looking at the screenshots, it's hard to find anything different without pixel peeping. Some of the stuff like the dirt on the ground is a bit blurrier but that's all I could tell at a quick glance
Also should point out that taking a screenshot compresses the data again so iit's not going to be a good way to make the comparison
Remuxing mkv to mp4 fucks stuffs up because mkv timebase is asssssssssss