Encoder improvements #127

nihil-admirari · 2024-06-19T19:58:39Z

Images:

The code no longer passes the options encoders doesn't expect to see: AVIF gets only AVIF-specific parameters, WEBP – only WEBP-specific ones etc.
Added support for animated AVIFs.

AVIF encoder's cpu-used is set to 6 compared to the default of 1. On my machine, values below 6 become increasing slower without much of an improvement when it comes to file size.

FFmpeg CRF 15 (high quality) 937×400 still images: 3 MB per 10k images in exchange for 1.5 to 3 times slower encoding.

cpu-used	file size in bytes	encoding time
0	13776	18.535s
1	13870	11.344s
2	13800	11.296s
3	13725	6.772s
4	14046	5.515s
5	14017	4.839s
6	14177	4.433s
7	14177	4.495s
8	14177	4.479s

FFmpeg CRF 32 (medium quality) 937×400 2 seconds animations: 1 MB per 100 images is more noticable, but the encoding time becomes obscene very very quickly.

cpu-used	file size in bytes	encoding time
0	84534	10m18.072s
1	85875	3m30.353s
2	89661	1m18.690s
3	90883	30.647s
4	91230	21.556s
5	90856	12.933s
6	92153	10.921s
7	92153	10.892s
8	92153	10.882s

Still image scaling now uses sinc, which is the best downscaling algorithm: https://stackoverflow.com/a/6171860. Animated images still use Lanczos.
Image scaling now uses accurate rounding.
min()s in scaling filter prevent image upscaling.
JPEG's quality is now correctly mapped into the range 2–31.
Image aspect ratio preserving scaling now uses -1 instead of -2. According to FFmpeg docs, -n scale is here to ‘make sure that the calculated dimension is divisible by n.’ I don't quite understand while divisibility by 2 was needed in the first place.

Audio:

Opus phase inversion is disabled, which improves mono quality: https://ffmpeg.org/ffmpeg-codecs.html#Option-Mapping.
The point of Opus is Ogg is lost on me, considering that Opus has its own extension. Ogg files containing non-Vorbis data are confusing. The ability to choose an Opus container was added to fix this, though Ogg is still the default.
Implementation of Opus in CAF was fixed. Previously it produced CAF-encoded files with Ogg extension.
Support for M4A and Webm containers have been added as a CAF replacement for newer Apple devices.
MP3-encoder now uses ABR, which is recommended for low bitrates: https://wiki.hydrogenaud.io/index.php?title=LAME#Recommended_encoder_settings.
loudnorm now has a proper two-pass implementation. dual_mono is now enabled to account for mono downmixing: https://ffmpeg.org/ffmpeg-filters.html#loudnorm.

Miscellanea:

append_user_audio_args no longer moves the last argument on every insert.
Exit codes of FFmpeg and MPV are now checked before reporting a successful conversion. Checking for file presence is not enough: sometimes failed conversions result in zero-length files.
Common MPV calling code was extracted into a separate function.
MPV's libaom-av1 support is now checked.

tatsumoto-ren · 2024-06-19T21:11:52Z

Hello. Thanks for the patch.

tatsumoto-ren · 2024-06-19T21:25:44Z

AVIF encoder's cpu-used is set to 6 compared to the default of 1. On my machine, values below 6 become increasing slower without much of an improvement when it comes to file size.

If a computer has less cores than 6, will it automatically choose the right number? Or would it be better to set cpu-used to min(6, nproc)?

tatsumoto-ren · 2024-06-19T21:38:10Z

Image aspect ratio preserving scaling now uses -1 instead of -2. According to FFmpeg docs, -n scale is here to ‘make sure that the calculated dimension is divisible by n.’ I don't quite understand while divisibility by 2 was needed in the first place.

Setting the value to -2 is required to ensure that encoding doesn't fail. If set to -1, there might be problems. Encoding will fail with an error like "height/width not divisible by 2". This is probably not something that happens often though.

tatsumoto-ren · 2024-06-19T22:36:38Z

.github/RELEASE/subs2srs.conf

+# Perform loudness normalization.
+# Parameter explanation can be found e.g. at:
+# https://auphonic.com/blog/2013/01/07/loudness-targets-mobile-audio-podcasts-radio-tv/
+# https://auphonic.com/blog/2019/08/19/dynamic-range-processing/
+loudnorm=yes
+loudnorm_target=-16
+loudnorm_range=11
+loudnorm_peak=-1.5


With the new config parameters, the users will have to update their config files. For example, I have previously set loudnorm parameters using ffmpeg_audio_args, and now I get audio files with different loudness.

Disabled two-pass loudnorm by default and brought back the custom filter. Added scary warnings that filter must be removed before enabling two-pass loudnorm.

Loudness will be the same if target, range and peak are set to the same parameters that previously were in ffmpeg_audio_args.

tatsumoto-ren · 2024-06-19T22:36:56Z

cfg_mgr.lua

- -- Animated webp images can only have .webp extension.
- -- The user has no choice on this.


Lost the comment here.

I thought this comment meant that there is only one animation format, compared to three formats for still pictures. Since I've added support for another animation format, I've removed the comment.

If the comment was actually supposed to state that WebP animations come with and only with .webp extensions, well, it's kinda obvious.

tatsumoto-ren · 2024-06-19T22:39:50Z

encoder.lua

+local function rescale_quality(quality, min_q, max_q)
+ local scaled = min_q + (max_q - min_q) * quality / 100
+ -- Round to the nearest integer that's better in quality.
+ if min_q > max_q then
+ return math.floor(scaled)
+ end
+ return math.ceil(scaled)


This function was previously named quality_to_crf but with the new name it's hard to understand what it does.

Renamed to fit_quality_percentage_to_range. JPEG function was also renamed since qscale is not CRF: https://stackoverflow.com/a/40673522. AVIF and JPEF constants were moved inside their respective functions.

tatsumoto-ren · 2024-06-19T22:42:11Z

encoder.lua

+ '-vf', string.format("scale='min(%d,iw)':'min(%d,ih)':flags=sinc+accurate_rnd",
+ self.config.snapshot_width, self.config.snapshot_height),
+ '-frames:v', '1',
+ table.unpack(encoder_args)


Using table.unpack() crashes mpvacious.

attempt to call field 'unpack' (a nil value)

Use h.unpack() instead.

Better be commented somewhere.

use h.unpack() to avoid crashing mpvacious

encoder.lua

nihil-admirari · 2024-06-20T07:15:08Z

If a computer has less cores than 6, will it automatically choose the right number? Or would it be better to set cpu-used to min(6, nproc)?

cpu-used has nothing to do with multithreading. It's a computational intensity parameter, kinda like compression_level for WebP. See https://ffmpeg.org/ffmpeg-codecs.html#Options-28:

cpu-used
Set the quality/encoding speed tradeoff. Valid range is from 0 to 8, higher numbers indicating greater speed and lower quality. The default value is 1, which will be slow and high quality.

tatsumoto-ren · 2024-06-20T07:20:19Z

encoder.lua

 end

-mpv.make_audio_args = function(source_path, output_path, start_timestamp, end_timestamp)
+mpv.make_audio_args = function(source_path, output_path,
+ start_timestamp, end_timestamp, args_consumer)


I wonder if it's possible to unwind these callbacks (e.g. args_consumer) to make the code more readable. Especially considering that lua doesn't have static typing.

There are two possible ways to do that:

Replace mp.command_native_async with mp.command_native. Can block the execution for considerable time (see the timings in the head), but may not be a problem if the rest of MPV remains responsive. Have no idea how MPV handles a single blocked Lua script.

Rewrite using Lua coroutines. Never worked with these.

What I meant is you could make the args table first, and then execute the args without passing the function that executes them into the function that makes the args.

tatsumoto-ren · 2024-06-20T07:30:15Z

cpu-used has nothing to do with multithreading. It's a computational intensity parameter, kinda like compression_level for WebP.

Alright. The name is confusing apparently.

…y and scale functions

nihil-admirari · 2024-06-20T14:32:25Z

Fixed user supplied filters not being appended after loudnorm.
Refactored the use of scaling related functions.

Encoder improvements

7e2391b

preserve aspect ratio using -2

0b31854

tatsumoto-ren reviewed Jun 19, 2024

View reviewed changes

tatsumoto-ren added 2 commits June 20, 2024 01:37

add comment

deb2391

restore lost comment

16755c4

tatsumoto-ren reviewed Jun 19, 2024

View reviewed changes

tatsumoto-ren added 4 commits June 20, 2024 01:43

update comment

0cc3671

format code

801ba2b

use h.unpack() to avoid crashing mpvacious

add comment

fe9a95a

add function quality_to_crf_avif()

1c47e59

tatsumoto-ren reviewed Jun 20, 2024

View reviewed changes

encoder.lua Outdated Show resolved Hide resolved

tatsumoto-ren reviewed Jun 20, 2024

View reviewed changes

encoder.lua Outdated Show resolved Hide resolved

tatsumoto-ren added 4 commits June 20, 2024 05:55

extract function

6df31e6

extract function

15a3be4

remove duplicate numbers

e252f91

add function quality_to_crf_jpeg()

665e44b

tatsumoto-ren reviewed Jun 20, 2024

View reviewed changes

Disable loudnorm for compatibility; user filters fix; refactor qualit…

0a7d1f6

…y and scale functions

Mention opus_container in README

c4be914

tatsumoto-ren merged commit aa86597 into Ajatt-Tools:master Jun 22, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Encoder improvements #127

Encoder improvements #127

nihil-admirari commented Jun 19, 2024

tatsumoto-ren commented Jun 19, 2024

tatsumoto-ren commented Jun 19, 2024

tatsumoto-ren commented Jun 19, 2024 •

edited

Loading

tatsumoto-ren Jun 19, 2024

nihil-admirari Jun 20, 2024

tatsumoto-ren Jun 19, 2024

nihil-admirari Jun 20, 2024

tatsumoto-ren Jun 19, 2024

nihil-admirari Jun 20, 2024 •

edited

Loading

tatsumoto-ren Jun 19, 2024

nihil-admirari Jun 20, 2024

nihil-admirari commented Jun 20, 2024

tatsumoto-ren Jun 20, 2024

nihil-admirari Jun 20, 2024

tatsumoto-ren Jun 21, 2024

tatsumoto-ren commented Jun 20, 2024

nihil-admirari commented Jun 20, 2024

		-- Animated webp images can only have .webp extension.
		-- The user has no choice on this.

Encoder improvements #127

Encoder improvements #127

Conversation

nihil-admirari commented Jun 19, 2024

tatsumoto-ren commented Jun 19, 2024

tatsumoto-ren commented Jun 19, 2024

tatsumoto-ren commented Jun 19, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

nihil-admirari Jun 20, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

nihil-admirari commented Jun 20, 2024

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

tatsumoto-ren commented Jun 20, 2024

nihil-admirari commented Jun 20, 2024

tatsumoto-ren commented Jun 19, 2024 •

edited

Loading

nihil-admirari Jun 20, 2024 •

edited

Loading