Some x264 and SVT-AV1 comparisons that don't make much sense to me

[asw3]

So with HanBrake 1.6.0 having the SVT-AV1 encoder I decided to do some testing.

From ffmpeg's wiki AV1 page:
"A CRF value of 23 yields a quality level corresponding to CRF 19 for x264, which would be considered visually lossless."

So I did some encodes.

The VMAF in these looks promising, but once you take a look at actual comparison screenshots you see that the AV1 encodes are significantly lower quality. Most easily seen in Spider-Man's costume (the lines that run through it): https://anonfiles.com/s7V8Z7O9y7/Test_7z

So what am I missing? Even at Preset 5 which would take forever to do an encode on the quality is still not the same as x264 crf 19 as is the claim on the wiki. I assume some CLI parameters must be set. Anything I am missing that can make a decent looking encode at presets 10-9 or maybe 8 that will save space and at least not look worse off? It seems SVT-AV1 really losses clarity and crispness.

[jagabo]

The equivalency is a general guideline. It may work for some videos, but not for all. In general, higher compression codecs get a lot of their additional compression by discarding or blurring away small, low contrast details -- like the grey lines on spiderman's suit.

[asw3]

Any way to reduce the blurring? Or essential parameters to set/consider with AV1?
As it stands now while the decrease in filesize on slower presets by itself looks very nice, it's definitely not worth the time needed to encode by itself. If we consider the lowered quality - why even use AV1?

[poisondeathray]

There are problems with all metrics, they are just a guide and the trends are more important

AV1 is work in progress (all encoders had their "growing pains" stages)

There tends to be oversmoothing with AV1, errors on scenechanges, the more motion the worse the problems - it's mainly due to the temporal filtering. It's a known issue

You can disable temporal filtering with --enable-tf 0 . The "penalty" is about 4-8% worse quality at a given bitrate . If using CRF encoding , the bitrate will change. The temporal filtering can be useful for low to very low bitrate scenarios

In general, the bitrate savings for av1 vs. x264 is larger at higher resolutions (same with x264 vs. x265; eg. the difference at UHD will be much larger at HD resolution)

[asw3]

Thank you for the suggestions and clarifications.

So for 1080p content, is x264 considered to be the best choice still?

[JN-]

I would use a present of medium for say x265 Av1 etc, it’s a good starting point. Why not then do some comparison, ssim, psnr, vmaf, of x264, vs av1.

Also pick a single data rate or output file size and use that target file size. Then see what the render quality metrics compare.

I hope to do that soon also, but comparing Nvenc Av1 HW to say Av1 Cpu, vs x264/hevc using a target output file size, all FHD. Then the RQM values will give some idea, but for sure eyeballing the results might be also necessary.

In any of this testing I only ever use a less than 30 second clip (includes people, water, foliage, indoor etc) for sanitys sake given how slow Av1 Cpu encoding can be.

[JN-]

Item no. Output files SSIM "All" SSIM "All" DB PSNR "Avg." VMAF Render Data rate Size MB Percentage %
time Aggregate

#1 Source-[CPU-AV1-Q31-'PR6 Medium'] 0.975079 16.034356 39.383845 96.696485 71s 10,720 33.6 100.00% #1
#2 Source-[NVENC-AV1-Q28-Medium] 0.968039 14.953740 38.300827 95.369925 2s 10,568 34.6 97.10% #2
#3 Source-[CPU-h264-Q25.6-Medium] 0.969114 15.102453 37.788991 94.553763 7s 10,214 33.4 96.83% #3
#4 Source-[CPU-hevc-Q27.1-Medium] 0.966272 14.720118 37.947905 94.776435 24s 10,237 33.5 96.32% #4
#5 Source-[NVENC-hevc-Q26-Medium]s 0.961199 14.111611 37.405059 94.098759 2s 10,261 33.6 94.72% #5
#6 Source-[QSV-hevc-Q27-Medium] 0.961173 14.108698 37.146669 93.929920 9s 10,795 35.3 94.51% #6
#7 Source-[NVENC-h264-Q26-Medium] 0.958872 13.858596 37.224572 94.114368 2s 10,550 34.5 94.15% #7
#8 Source-[QSV-h264-Q33-Medium] 0.958450 13.814251 36.975956 92.508732 4s 10,093 33.1 93.50% #8

[asw3]

You can disable temporal filtering with --enable-tf 0.

I tried doing a HandBrake encode with enable-tf=0 in the Advanced Options but I don't think it did anything. Is that how it's supposed to be written down for HandBrake, or is the syntax different? I can't find any documentation about this.

[jagabo]

I tried doing a HandBrake encode with enable-tf=0 in the Advanced Options but I don't think it did anything. Is that how it's supposed to be written down for HandBrake, or is the syntax different?

That's the normal syntax for a single option in the extra options box with other codecs (no leading - or --). I don't know if the av1 interface supports the option.

[poisondeathray]

Not sure what it is in handbrake; --enable-tf 0 is for SvtAv1EncApp . In ffmpeg it would be -svtav1-params enable-tf=0

[asw3]

HandBrake uses it's own denomination style for x264 and x265, so I assume it's the same for AV1. For example x264

--bframes 4 --ref 4

in HandBrake should be written as

bframes=4:ref=4

HandBrake doesn't seem to accept neither -- nor it's usual denomination style for AV1.

[JN-]

I updated my previous post with AV1 Nvenc HW encoding comparison. I would much prefer to post an image, say jpg, I cannot see how to do it. Ok, managed it.




[Attachment 72378 - Click to enlarge]

No B frames were used/harmed in any way.

[JN-]

I couldn't include a FHD AMD comparison as it always outputs 1082 instead of 1080. So I did the AV1 in UHD. Hevc comparison also included for CPU and Nvenc encoding.



[Attachment 73622 - Click to enlarge]