Which of these 2 looks better?

[sophisticles]

I was doing a few encoding tests comparing the Ice Lake qsv h264 encoder vs x264 and I got these interesting results when I used vmaf, psnr and ms-ssim to compare the quality.

If I do a frame by frame comparison, qsv scores higher in 1251 frames compared to 104 frames in which x264 scores higher, the rest are a tie at a score of 100, when compared with vmaf.

When compared with psnr and ms-ssim, it's a knockout win for x264 in both metrics, with x264 having higher scores in all 2845 frames.

I would like some opinions as to which results your guys agree with.

The source is too large to attach here, but you will recognize the test samples from Xiph, I took the 1080p25 samples. and created one source file encoded lossless UT.

[poisondeathray]

I was doing a few encoding tests comparing the Ice Lake qsv h264 encoder vs x264 and I got these interesting results when I used vmaf, psnr and ms-ssim to compare the quality.

Is the question "what looks better ?" as in subjectively? or "what looks more like the source ?"

qsv encode has some enhancement like oversharpening , local contrast enhancement applied . Subjectively I think it would look better if it was tuned down, because right now it just enhances artifacts and ringing and is currently applied way to strongly

If I do a frame by frame comparison, qsv scores higher in 1251 frames compared to 104 frames in which x264 scores higher, the rest are a tie at a score of 100, when compared with vmaf.

Which vmaf mode ?

One of the "tricks" of scoring higher in original vmaf was applying sharpening if you read the vmaf blog . There were complaints that it would abused

The VMAF Neg mode was introduced to account for this - thus sharpening, tune vmaf or histogram equilization tend to score lower now with vmaf neg, instead of higher with vmaf original

https://netflixtechblog.com/toward-a-better-quality-metric-for-the-video-community-7ed94e752a30

[rrats]

well x264 doesn't add sharpening and has a bit less artefacts so theres that

[_Al_]

That sharpened thing looks definitely better when watched as a clip, but it is totally ruined for other work. It is just final thing to watch, where things "look clearer" even if there is not much more details because it is sharpened. When posting things on web etc., it is a strategy to sharpen and also enhance colors a bit, especially for images. Folks just tell you it is better. But you cannot come up back to those things and work on them again.

You do not want to see details, for example even 640x360 cut out of that bee, noticing black halos and how patterns are not smooth anymore, how everything looks broken. But just watching it in a clip, full resolution, it looks better. So the answer, better is that sharpened version, but as soon it is out there, it is good for nothing further, it is for watching only.
Click on each images three times to have it in a tab for comparisons, click on each for three times to have a full blow up in a separate TAB.

[Attachment 70064 - Click to enlarge]

[Attachment 70065 - Click to enlarge]

[sophisticles]

Thanks for the feedback, my assessment was that x264 was a closer match to the source, meaning the psnr and ms-ssim scores where right but that the qsv encode looked better, more pleasing, meaning that the vmaf score was right in that regard, seeing how vmaf is supposed to better correlate with people's perceptions.

QSV has a vpp filter that is called "detail-enhancement", that seems like it's a sharpening and contrast adjustment filter and it's runs fully on the fixed function unit.

If you were doing just one encode, then you could create a custom script with a custom filter chain, crank up the x264 settings and wait until it finished, but assume you are going to do hundreds, or thousands of encodes, like say you are converting your collection of Buffy, Angel, Honeymooners, Supernatural and various movie dvd and b;u-rays to a format that you can put on a thumb drive and take with you on trips easily.

Then time, heat generated and electricity costs become a factor and qsv becomes a very appealing alternative.

Attached are some more test encodes, with the vpp filter turned down.

With qsv the highest quality encode is using cqp, but i also did some using la-icq because Intel has said that this is the closest mode to crf that they have.

As the filter strength was turned down I had to adjust cqp and la-icq to maintain the bit rate as close as possible.

[sophisticles]

Here's one I did with just fixed function, that means no Lookahead and no B frames.

Just realized that there is no attachment.