Intel SVT-AV1

[Blue_MiSfit]

I've yet to try either! Good to know, I'll see if I can set some time aside for analysis.

[quietvoid]

SVT-AV1 2 pass still doesn't work properly with a higher target bitrate. I set -tbr for 10 mbps and the resulting encode is only 4-5 mbps.
Both -rc 2 (4.6 mbps) and -rc 3 (3.8 mbps).

[Blue_MiSfit]

Hm, it seemed to work fine in my 1080p smoke test clip. I got a 10 Mbps output in 2 pass mode using -rc 2

[Blue_MiSfit]

I get some really bad dealbreaker artifacts at high bitrates using fast presets though. For example, my smoke test at QP 30 with enc-mode 8 through 5 has really bad glitches in certain parts of the frame. These should not be happening at 12 Mbps

They clear up at enc-mode 4

https://imgur.com/a/46Oeofq

[quietvoid]

Originally Posted by Blue_MiSfit

Hm, it seemed to work fine in my 1080p smoke test clip. I got a 10 Mbps output in 2 pass mode using -rc 2

I've exhausted all the options I could think of, -rc 2 -tbr 10000000 never gives me a bitrate close to the target.

Trimmed 500 frames of a BT.709 1080p 8-bit SDR clip, ran these two lines:

SVT-AV1\SvtAv1EncApp.exe -i sample.y4m -fps-num 24000 -fps-denom 1001 -rc 2 -tbr 10000000 -output-stat-file out.stats -enc-mode-2p 4 -b out.ivf

Result: 716.24 kbps

SVT-AV1\SvtAv1EncApp.exe -i sample.y4m -fps-num 24000 -fps-denom 1001 -rc 2 -tbr 10000000 -input-stat-file out.stats -enc-mode 4 -b out.ivf

Result: 455.21 kbps

[Blue_MiSfit]

Maybe some weird edge case with y4m input? I've been doing my tests with raw yuv.

Try this exact 2 pass command (my documentation PR just got merged )

https://github.com/OpenVisualCloud/S...-command-lines

[Tadanobu]

2 pass will only work for me if I specify the number of frames to encode (-n). Otherwise 1st pass works fine but 2nd pass freezes at first frame. But I've noticed I've had the following error several times : "Error in freed returnVal 0"

[quietvoid]

Originally Posted by Blue_MiSfit

Maybe some weird edge case with y4m input? I've been doing my tests with raw yuv.

Try this exact 2 pass command (my documentation PR just got merged )

https://github.com/OpenVisualCloud/S...-command-lines

Raw .yuv was an option tested
I've tried again with the example CLI and it behaves the same.
I'll just give up, I don't think it's implemented properly yet.

I had asked 4 months ago on another site and got this response:

Target bitrate only works in mode 1, and as you said, It's not implemented yet. The target_bitrate variable is for future use and currently is ignored.

Mode 1 refers to the rate control mode, and that mode is still disabled.

[Blue_MiSfit]

Any chance you can share the sample? I'm sure the developers would be interested by that.

If you use constant QP can you achieve a higher bitrate?

[quietvoid]

Yes, using -q does change the bitrate. It's only -tbr not reaching the target bitrate.

The sample isn't special, I've tried with 24 frames scenes, 500 frames of one scene and multiple samples.

I'm sure I could reproduce it with Tears of Steel but I'll wait for a couple PRs to be merged first (including the kbps one )

-tbr has never worked for me, 2 pass still hasn't changed that.

[NikosD]

Ryzen 3950X is very fast on SVT-AV1, no doubt about it, it's even faster than latest HEDT from Intel Core i9-10980XE which has 2 more cores, BUT Threadripper's devastating performance is simply out of this world.

It's three times faster than previous Threadripper 32C and two times faster than latest Intel 18C

[Funky080900]

100 fps using -enc-mode 4?

[birdie]

SVT-AV1 0.75 is out with a long list of changes.

I've compared it to x265 at a similar bitrate and x265 came up on top - a lot more details were generally preserved and SVT-AV1 output is kinda blurry/too smooth.

ffmpeg -i source.mkv -vf fps=30 -an -c:a copy -c:v libx265 -preset veryslow \
-x265-params keyint=600:min-keyint=30:bframes=16:crf=20:no-sao=1 265.mkv
ffmpeg -hide_banner -loglevel panic -nostats -i source.mkv -nostdin -f rawvideo -pix_fmt yuv420p - \
| ./SvtAv1EncApp -i stdin -w 1280 -h 720 -n 25038 -b out.av1 -enc-mode 1 -q 24 -rc 0 -intra-period 250

The encoding speed of SVT-AV1 was roughly 0.2fps and x265 ran at around 12fps. x265 compressed to 822.96 kb/s, svt-av1 to roughly 750 kb/s.

Maybe there are flags to retain details better but I didn't find them.

enc-mode 0 is even slower and basically impossible to use.

[RanmaCanada]

Originally Posted by Funky080900

100 fps using -enc-mode 4?

HAHA, no. Not even close

[osgZach]

I'm actually more shocked to see a performance graph that actually uses an FPS metric than anything. I got frustrated and stopped looking at such charts because they always had useless, or obscure metrics that involve extra math to find out what the performance is; and that's on top of them not even telling you the encode settings or other critical information


Someone should ask Toms what the hell "1080p encoder mode 4 to AV1 encode" even means and if they speak english.

[quietvoid]

Version 0.8.0, still getting final bitrates way off of the target I set.

SVT-AV1\SvtAv1EncApp.exe -i tos_10b.y4m -fps-num 24000 -fps-denom 1000 -bit-depth 10 -intra-period 24 -irefresh-type 2 -rc 1 -tbr 10000 -tile-rows 3 -tile-columns 3 -output-stat-file out.stats -enc-mode-2p 4 -b out.ivf
SVT-AV1\SvtAv1EncApp.exe -i tos_10b.y4m -fps-num 24000 -fps-denom 1000 -bit-depth 10 -intra-period 24 -irefresh-type 2 -rc 1 -tbr 10000 -tile-rows 3 -tile-columns 3 -input-stat-file out.stats -enc-mode 4 -b out.ivf

Resulting encode: 6305.63 kbps

[NikosD]

Running Linux, tested by Phoronix.
https://i.postimg.cc/nL6c7W9H/svt-av1-39xx-9900.png

[foxyshadis]

SVT-AV1 0.8.1 has been released. Mostly a bugfix release, plus 10-bit palette support for screen coding.

[utack]

Does anyone know how the SVT-AV1 Gui does 2pass?
2 times the same speed or first pass at speed 8?

[foxyshadis]

Originally Posted by utack

Does anyone know how the SVT-AV1 Gui does 2pass?
2 times the same speed or first pass at speed 8?

It uses the same speed on both passes, no fast-first-pass. That doesn't seem ideal, the consensus is that speed 7 or 8 is fine for generating statistics.