x265 HEVC Encoder

[Boulder]

Originally Posted by vpupkind

Have you tried higher threshold values? The algorithm does not look at motion, only on colors and texture. They differ a lot when you have a strong scene cut, but low thresholds can result in it false positives.

With the default threshold, lots of false positives. Min-keyint is respected as all were i-frames. The only I-frame was right at the beginning of the file, the same in the encode with the standard scenecut detection. Due to open GOP?

Some tests; Hot Fuzz is an easy clip with clear scene cuts. The correct amount of keyframes in the testclip is 13.

Threshold 0.5, it misses all scene changes
Threshold 0.1, it misses a lot of them
Threshold 0.05, it still misses some quite obvious ones
Threshold 0.025, it begins to place too many (29)
Threshold 0.03, still too many (19)
Threshold 0.035, almost the correct amount (14)

The default encoding missed two very obvious scene cuts which the histogram one was able to find. The histogram one had one false detection in a scene of higher motion. The false one was in the middle of the scene and slightly less blurry, and I suspect that red text on a white shirt skewed the detection routine there even more.

[Boulder]

I uploaded the source clip here: https://drive.google.com/open?id=19c...J08OaQbfBIHUjm

To encode, I cropped off the black borders and downscaled with Spline36 to 1280x544. The histogram sc detection picks up frame 805 incorrectly, others seem good to me.
For some reason, I am unable to tweak the --scenecut and --scenecut-bias parameters to get the standard sc detection to pick up frames 380 and 582 which are perfect cuts between scenes.

[foxyshadis]

Originally Posted by Boulder

I uploaded the source clip here: https://drive.google.com/open?id=19c...J08OaQbfBIHUjm

To encode, I cropped off the black borders and downscaled with Spline36 to 1280x544. The histogram sc detection picks up frame 805 incorrectly, others seem good to me.
For some reason, I am unable to tweak the --scenecut and --scenecut-bias parameters to get the standard sc detection to pick up frames 380 and 582 which are perfect cuts between scenes.

At this point, copy-pasting everything you've discovered into the x265 bug tracker sounds like the way to go.

[Boulder]

Originally Posted by foxyshadis

At this point, copy-pasting everything you've discovered into the x265 bug tracker sounds like the way to go.

Agreed, I'll do that right away.

[Stereodude]

Are there no more Wolfberry x265 builds (since mid 2019) or another x265 build that can open a .avs directly without the need for piping?

[Forteen88]

I'd appreciate it if devs here, that releases x265-builds, comments on what compiling-setting they use!
My favorite is compiler-setting is "Optimization: Compile for speed", although maybe it doesn't matter that much since x265 is mostly coded in assembler.

[Stereodude]

Originally Posted by HolyWu

My build is compiled with lavf support.

Thanks for pointing that out. I had missed that detail. I hate having to pipe into x265.

[fauxreaper]

Is it viable to make a --frame-dup mode 2 without generating VFR video? I don't know why x265 spends so many bits on duplicated frames.

Example:

1) Download wallpaper on https://wallpaperscraft.com/download...4996/1920x1080
2) Create avs with
ImageSource("planet_clouds_light_star_94996_1920x1080.jpg",start=0,end=199,fps=23.976)
3) Encode avs with x265 x64 10bit, crf 22, preset slow and --csv-log-level 2 --csv
4) Open csv and examine used bits

https://i.imgur.com/vSwL4KQ.png

Why frames with 100% skip 64x64 need to use more than one thousand bits? Why the bits variation on duplicated frames, shouldn't almost all of them have similar size in bits?

x264 has much less variation on bitrate with same video. x64, 10bit, crf22.

I've filed an issue on x265's bugtracker, https://bitbucket.org/multicoreware/...t-variation-on

[jlpsvk]

change from 3.3+1 to 3.3.+2 that haven't been mention anywhere??? command give "unknown parameter" with 3.3+2

--no-rskip --> --rskip 0

[benwaggoner]

Originally Posted by fauxreaper

Is it viable to make a --frame-dup mode 2 without generating VFR video? I don't know why x265 spends so many bits on duplicated frames.

What is mode 2? The syntax just has --frame-dup as a Boolean.

Frame-dup should be fully compatible with constant frame rate, since it is putting in skip frames, not changing frame durations (which was how WMV9 did it back in the day).

[Greenhorn]

Originally Posted by jlpsvk

change from 3.3+1 to 3.3.+2 that haven't been mention anywhere??? command give "unknown parameter" with 3.3+2

--no-rskip --> --rskip 0

There's information at (https://x265.readthedocs.io/en/default/cli.html) (just search for "rskip"), or if you run x265 --fullhelp from the CLI.

Basically, --rskip 0 is (old) --no-rskip, (old) --rskip 1 is --rskip, --rskip 2 and --rskip 3 use a different method of analysis to decide when to perform the skip. --rskip-edge-threshold controls how much recursion the encoder skips, if you've set --rskip 2 or 3.

From my limited testing it seems to still need

[jlpsvk]

Originally Posted by Greenhorn

There's information at (https://x265.readthedocs.io/en/default/cli.html) (just search for "rskip"), or if you run x265 --fullhelp from the CLI.

Basically, --rskip 0 is (old) --no-rskip, (old) --rskip 1 is --rskip, --rskip 2 and --rskip 3 use a different method of analysis to decide when to perform the skip. --rskip-edge-threshold controls how much recursion the encoder skips, if you've set --rskip 2 or 3.

From my limited testing it seems to still need

In the docs, there it is, but nowhere in changelog... That's what i am talkong about.. I am using --rskip 0 (--no-rskip till now)...

[stax76]

RD Level    Rskip Mode    Skip Recursion Heuristic

0 - 4       1             Neighbour costs.
5 - 6       1             Comparison with inter2Nx2N.
0 - 6       2             CU edge denstiy.
0 - 6       3             CU edge denstiy with forceful skip for lower levels of CTU.

Is row 1 column 2 correct?

What is the default? --rskip 0 is the same as no --rskip ?

https://x265.readthedocs.io/en/defau...mdoption-rskip

I hope I added it correctly.

New OptionParam With {
    .Switch = "--rskip",
    .Text = "Recursion Skip",
    .Expand = True,
    .Options = {
        "Disabled",
        "0 - RD Level 0-4 Neighbour costs",
        "1 - RD Level 5-6 Comparison with inter2Nx2N",
        "2 - RD Level 0-6 CU edge denstiy",
        "3 - RD Level 0-6 CU edge denstiy with forceful skip for lower levels of CTU"},
    .Values = {"", "0", "1", "2", "3"}}

[jlpsvk]

@Stax76

--rskip 0 should be DISABLED (should do the same as --no-rskip) according to docs

[Boulder]

So --rskip 1 is now the same as --rskip previously?
Once again, would be very nice to know which use cases would benefit from the other settings. To me it looks like the changes commited lately have something to do with cartoons and anime, which I never touch myself.

[nevcairiel]

Originally Posted by stax76

RD Level    Rskip Mode    Skip Recursion Heuristic

0 - 4       1             Neighbour costs.
5 - 6       1             Comparison with inter2Nx2N.
0 - 6       2             CU edge denstiy.
0 - 6       3             CU edge denstiy with forceful skip for lower levels of CTU.

Is row 1 column 2 correct?

It is correct. "--rskip 1" behavior depends on the RD Level, it differs between RD Level 0-4 and 5-6.

So, your option mapping would not be correct. 0 is disabled. 1 is both those modes, depending on RD.
Default is actually 1, and not disabled, unless --tune grain is used, according to the docs.

[RainyDog]

Originally Posted by Boulder

So --rskip 1 is now the same as --rskip previously?

Appears so, yes. I've not changed my command line since the change and it just defaults to --rskip 1 now.

I'll have a play around with --rskip 2 and 3 shortly but looks like there's also a --rskip-edge-threshold setting to take into account with those too.

I didn't find the speed difference between the old --rskip and --no-skip to be worth the trade off though, so unless the speed drop-off with --rskip 2 and 3 is substantially less then I don't expect them to be of much use.

[stax76]

@nev

I think I got it now, thank you.

New OptionParam With {
    .Switch = "--rskip",
    .Text = "Recursion Skip",
    .Expand = True,
    .Options = {
        "Undefined",
         "0 - Disabled",
         "1 - RD Level 0-4 Neighbour costs, RD Level 5-6 Comparison with inter2Nx2N",
         "2 - RD Level 0-6 CU edge denstiy",
         "3 - RD Level 0-6 CU edge denstiy with forceful skip for lower levels of CTU"},
    .Values = {"", "0", "1", "2", "3"}},

[foxyshadis]

Originally Posted by jlpsvk

In the docs, there it is, but nowhere in changelog... That's what i am talkong about.. I am using --rskip 0 (--no-rskip till now)...

If you want to use the bleeding edge builds, you have to be prepared for things to break occasionally, and UIs to catch up later. Otherwise, stick to stable builds or the build provided by the UI.

[MeteorRain]

Originally Posted by jlpsvk

In the docs, there it is, but nowhere in changelog... That's what i am talkong about.. I am using --rskip 0 (--no-rskip till now)...

I just checked the commit log, and found that rskip changes is on future 3.4. You must have used a 3.3+X where X includes the 3.4 changes.

It could be confusing to write 3.3+X in version number indicating a future 3.4 dev version.