x265 HEVC Encoder

[filler56789]

Originally Posted by Wolfberry

Yeah, that build requires AVX.

New build

x265 [info]: HEVC encoder version 3.0_Au+21-bac0e1acb874
x265 [info]: build info [Windows][GCC 8.3.1][64 bit] 8bit+10bit+12bit
x265 [info]: (libavcodec  58.52.100)
x265 [info]: (libavformat 58.27.103)
x265 [info]: (libavutil   56.26.100)
x265 [info]: (lsmash       2.16.1)

I'm using march=core2 for this build (I use --with-arch=core2 when configuring gcc), so it should work for most people, but no guarantee as I don't have a spare machine to test.

Thanks for the compatible build

It is somewhat bloated because of the additional stuff, but I compressed it with upx and now it has a healthy weight.

[Boulder]

How does one get rid of the "onion gradient" artifacts of x265? In the sample images, they are apparent in the character's forehead and also in the purplish stripe near his hair (zoom a bit to see both better). The bitrate of the encodes is ~6900 kbps.

x264


x265, aq-mode 1


x265, aq-mode 2


Some base settings of the x265 encode:

--deblock -2:-1 --no-strong-intra-smoothing --cbqpoffs -3 --crqpoffs -3 --subme 3 --merange 24 --no-sao --no-rect --qcomp 0.7 --rd 6 --rd-refine
--aq-mode 1/aq-mode 2 --aq-strength 0.9 --ctu 16 --max-tu-size 8 --qg-size 16 --tu-inter-depth 2 --tu-intra-depth 2 --limit-tu 1 --limit-refs 3
--max-merge 2 --ref 5 --bframes 10

The x264 settings are pretty much --preset veryslow --tune film --no-fast-pskip --no-dct-decimate --deblock -2:-1 --qcomp 0.7.

[mini-moose]

Originally Posted by Boulder

How does one get rid of the "onion gradient" artifacts of x265? In the sample images, they are apparent in the character's forehead and also in the purplish stripe near his hair (zoom a bit to see both better).

Maybe post a source cap too.

[benwaggoner]

Originally Posted by Boulder

How does one get rid of the "onion gradient" artifacts of x265? In the sample images, they are apparent in the character's forehead and also in the purplish stripe near his hair (zoom a bit to see both better).

Some base settings of the x265 encode:

--deblock -2:-1 --no-strong-intra-smoothing --cbqpoffs -3 --crqpoffs -3 --subme 3 --merange 24 --no-sao --no-rect --qcomp 0.7 --rd 6 --rd-refine
--aq-mode 1/aq-mode 2 --aq-strength 0.9 --ctu 16 --max-tu-size 8 --qg-size 16 --tu-inter-depth 2 --tu-intra-depth 2 --limit-tu 1 --limit-refs 3
--max-merge 2 --ref 5 --bframes 10

The x264 settings are pretty much --preset veryslow --tune film --no-fast-pskip --no-dct-decimate --deblock -2:-1 --qcomp 0.7.

You are doing some weird settings in there which may have combinatorial effects. Can you try just --preset slower and see how that turns out?

A 8x8 max tu is pretty unusual, and those are also big chroma offsets which can suck bits away from luma. What are the goals of these settings, and how tdid you come to them?

[Boulder]

I got you the original clip here. I basically did a very slight MDegrain and then downscaled to 1280x720 using a very sharp Bicubic and fed to the encoders.

https://drive.google.com/open?id=1Qm...9NZavzXFTSkGIr

My x265 settings are based on frame-by-frame comparisons that I made several times. A small CTU and TU size look better (least distortion compared to the original image) with 720p encodes. For 1080p, I would use one notch higher values (also based on my tests). The chroma offsets don't change the bitrate much in CRF mode so the difference in 2-pass shouldn't also be big. Besides, x264 already does a similar thing by default.

EDIT: does anyone else have problems with topic notifications? I don't get any emails at all from any thread I've subscribed to.

[excellentswordfight]

Originally Posted by Boulder

I got you the original clip here. I basically did a very slight MDegrain and then downscaled to 1280x720 using a very sharp Bicubic and fed to the encoders.

https://drive.google.com/open?id=1Qm...9NZavzXFTSkGIr

My x265 settings are based on frame-by-frame comparisons that I made several times. A small CTU and TU size look better (least distortion compared to the original image) with 720p encodes. For 1080p, I would use one notch higher values (also based on my tests). The chroma offsets don't change the bitrate much in CRF mode so the difference in 2-pass shouldn't also be big. Besides, x264 already does a similar thing by default.

EDIT: does anyone else have problems with topic notifications? I don't get any emails at all from any thread I've subscribed to.

Dont seem to be an "general" issue with x265. Did a re-encode of your sample and I didnt get an gradiant issue, and overall I think x265 did a better job. I think benwaggoner is right, its your settings thats causing it.

x264 2pass @ 5Mbps, veryslow tune film


x265 2pass @ 5Mbps, slow, --no-sao --no-strong-intra-smoothing --deblock -1:-1

[Boulder]

Originally Posted by excellentswordfight

Dont seem to be an "general" issue with x265. Did a re-encode of your sample and I didnt get an gradiant issue, and overall I think x265 did a better job. I think benwaggoner is right, its your settings thats causing it.

Thanks, that's exactly what I was looking for since I was experimenting a lot with aq-strength and the psy options to try to figure out what is wrong but was clearly looking in the wrong place.

I did some more testing now that I found those troublesome frames (it seems they evaded me when I pinpointed my go-to settings a long time ago), and using CTU 32 makes the onionize effect go away. I also left max-tu-size and qg-size at their defaults since things looked good.

[jethro]

Originally Posted by Boulder

Thanks, that's exactly what I was looking for since I was experimenting a lot with aq-strength and the psy options to try to figure out what is wrong but was clearly looking in the wrong place.

I did some more testing now that I found those troublesome frames (it seems they evaded me when I pinpointed my go-to settings a long time ago), and using CTU 32 makes the onionize effect go away. I also left max-tu-size and qg-size at their defaults since things looked good.

Encoder Settings:
x265 CRF 28, veryslow, --no-sao --no-strong-intra-smoothing --deblock -1:-1








More bitrate removes the artifacting.

[Boulder]

In those cases, I had thrown quite a lot of bitrate there. Using CRF18 produced a smaller file than 6900 kbps, which was x264@CRF18. CRF28 is something I'd never use The onion artifacts were visible only in certain frames with very high motion, that's probably why I missed them in my test rounds earlier. I do recall someone complaining that tune grain sometimes did those as well.

[Blue_MiSfit]

I'd caution against doing too much settings tuning using still frame comparisons, especially when evaluating max CU size (--ctu).

I've spent a lot of time doing this. Using --ctu 32 can indeed produce a sharper / more detailed / less blurred image when looking at single frames - relative to keeping the default (--ctu 64) in some cases. However, the trade-off is typically more blocking, which you're more likely to notice in motion. In other words, an inferior user experience unless you're using very high bitrates.

You're kind of performing a brute force non-adaptive psy tuning. x265 will use smaller CUs when it sees fit, you're just forcing it to never consider 64x64, even when coding giant uniform areas.

In my opinion, you're typically best off trusting the extensive tuning and testing that MulticoreWare has done, and just using a preset. They set --ctu to 64 starting with --preset veryfast, and keep it that way all the way through placebo. In fact, in looking at the documentation for this parameter: https://x265.readthedocs.io/en/defau...#cmdoption-ctu it appears the only reason they set this to 32 for superfast and lower is that this increases parallelism and can therefore be faster.

Do what's right for you, of course

[tuanden0]

I got crash when using zone in x265_12bit.
So I tried to use VSEdit to test it and here is result.
Here is x265 file: http://msystem.waw.pl/x265/x265-3.0_..._gcc83-AVX2.7z
Here is everything else: https://i.imgur.com/FFlFylW.png

[Boulder]

Originally Posted by Blue_MiSfit

I'd caution against doing too much settings tuning using still frame comparisons, especially when evaluating max CU size (--ctu).

I've spent a lot of time doing this. Using --ctu 32 can indeed produce a sharper / more detailed / less blurred image when looking at single frames - relative to keeping the default (--ctu 64) in some cases. However, the trade-off is typically more blocking, which you're more likely to notice in motion. In other words, an inferior user experience unless you're using very high bitrates.

You're kind of performing a brute force non-adaptive psy tuning. x265 will use smaller CUs when it sees fit, you're just forcing it to never consider 64x64, even when coding giant uniform areas.

In my opinion, you're typically best off trusting the extensive tuning and testing that MulticoreWare has done, and just using a preset. They set --ctu to 64 starting with --preset veryfast, and keep it that way all the way through placebo. In fact, in looking at the documentation for this parameter: https://x265.readthedocs.io/en/defau...#cmdoption-ctu it appears the only reason they set this to 32 for superfast and lower is that this increases parallelism and can therefore be faster.

Do what's right for you, of course

I've always considered 64x64 to be quite a lot for 720p video, and also the difference in performance is substantial. I can see that the average QP is lower with 32x32 than 16x16 though so it could be that there are situations where the encoder could squeeze some more out of the video. Of course, I could always run two simultaneous encodes to keep the CPU as busy as possible.

As you apparently have been testing things quite a lot, do you have any opinion regarding AQ-mode? I'm still quite unsure if the new default of mode 2 is the best general option for film.

[excellentswordfight]

Originally Posted by Boulder

I've always considered 64x64 to be quite a lot for 720p video, and also the difference in performance is substantial. I can see that the average QP is lower with 32x32 than 16x16 though so it could be that there are situations where the encoder could squeeze some more out of the video. Of course, I could always run two simultaneous encodes to keep the CPU as busy as possible.

As you apparently have been testing things quite a lot, do you have any opinion regarding AQ-mode? I'm still quite unsure if the new default of mode 2 is the best general option for film.

I wrote to Ma a fem months back regarding ctu 64 being overkill for reslutions under 2160p, and asking why this as the default value. And he responded with numbers that showed a loss in compression with using lower ctu values, even for lower res material. The difference is probably pretty negligible though, but i dont see any reason to lower it IF youre not trying to saturate more threads (were it can do wonders). I only lower it to 32 for 1080p if i’m at 16 threads or more.

[Barough]

x265 v3.0_Au+22-feec4bdf9866 (32 & 64-bit 8/10/12bit Multilib Windows Binaries) (32bit-GCC v7.4.0 / 64bit-GCC v8.3.0)

https://bitbucket.org/multicoreware/x265/commits/branch/default

[Blue_MiSfit]

The difference isn't huge, but it does help.

IMO aq-mode 2 is definitely better in most cases. I set this always unless doing extremely low bitrate when I'll use mode 3.

[~ VEGETA ~]

I wanted to encode a 10-bit H.264 MKV video (lossless, 30 gb) to HEVC 10-bit... I piped it via ffmpeg but I put this "-pix_fmt yuv420p" instead of this "yuv420p10le"... I suspect this is wrong, but it is the only way it worked. Can you kindly guide me here?

I encoded .vpy file to lossless 10-bit H.264 since my device is slow with the filters I used... so now how to encode the 10-bit .mkv using x265? << I assumed using -pix_fmt yuv420p is wrong.



Plus, next time when I want to encode like this (lossless 10-bit H.264 then 10-bit HEVC), should I encode to .yuv/.h264 in x264 (--muxer yuv\raw)? So I can input it directly to x265 without piping??



Here is my command line:

"C:\.....\ffmpeg.exe" -i "C:\....\lossless_H264_video.mkv" -pix_fmt yuv420p -f yuv4mpegpipe - | C:\....\x265_64bit_10bit.exe --y4m --preset slower --crf 17 --ref 6 --rd 6 --psy-rd 1 --ctu 64 --aq-strength 0.8 --output-depth 10 --input-res 1920x1080 - --output "C:\....\ep07_1080p_video_HEVC.hevc"

[sneaker_ger]

Post your complete command line and ffmpeg log.

(Yes, if you want 10 bit output setting -pix_fmt yuv420p is wrong. Make sure your ffmpeg build is compiled with 10 bit libx265.)

[~ VEGETA ~]

Originally Posted by sneaker_ger

Post your complete command line and ffmpeg log.

(Yes, if you want 10 bit output setting -pix_fmt yuv420p is wrong. Make sure your ffmpeg build is compiled with 10 bit libx265.)

I posted my command line as you see now.

Where can I find such ffmpeg build for windows?

I thought -pix_fmt is to get the input as 10-bit...

[sneaker_ger]

I thought you wanted to use libx265 from within ffmpeg. Ignore my last post.

Use:

ffmpeg -i "INPUT" -f yuv4mpegpipe -strict -1 -

Then make sure the log says it's "yuv420p10le".

[excellentswordfight]

Originally Posted by ~ VEGETA ~

I wanted to encode a 10-bit H.264 MKV video (lossless, 30 gb) to HEVC 10-bit... I piped it via ffmpeg but I put this "-pix_fmt yuv420p" instead of this "yuv420p10le"... I suspect this is wrong, but it is the only way it worked. Can you kindly guide me here?

I encoded .vpy file to lossless 10-bit H.264 since my device is slow with the filters I used... so now how to encode the 10-bit .mkv using x265? << I assumed using -pix_fmt yuv420p is wrong.



Plus, next time when I want to encode like this (lossless 10-bit H.264 then 10-bit HEVC), should I encode to .yuv/.h264 in x264 (--muxer yuv\raw)? So I can input it directly to x265 without piping??



Here is my command line:

"C:\.....\ffmpeg.exe" -i "C:\....\lossless_H264_video.mkv" -pix_fmt yuv420p -f yuv4mpegpipe - | C:\....\x265_64bit_10bit.exe --y4m --preset slower --crf 17 --ref 6 --rd 6 --psy-rd 1 --ctu 64 --aq-strength 0.8 --output-depth 10 --input-res 1920x1080 - --output "C:\....\ep07_1080p_video_HEVC.hevc"

I never had issues with the following:

"ffmpeg.exe" -i "input.mkv" -pix_fmt yuv420p10le -f yuv4mpegpipe -strict -1 - | "x265.exe" --y4m

So maybe you need to ad -strict -1