Auto-Boost-Essential

Auto-Boost-Essential

Auto-Boost-Essential is the latest iteration of the Auto-Boost formula, streamlined and refined for greater convenience, speeds and boosting gains!

This encoding script is intended to be paired with my SVT-AV1-Essential encoder fork.
SVT-AV1-Essential sports excellent quality consistency, but Auto-Boost-Essential offers exceptional consistency!

Here is how it works: the script runs a first encoder fast-pass, finds scenes based on the introduced keyframes, calculate metrics scores, automatically adjusts the CRF of scenes in order to increase quality consistency and then runs a final-pass with these adjustements.

The quality metrics at play this time again are SSIMULACRA2 and XPSNR.

Dependencies

Before using Auto-Boost-Essential, you will need to install the following dependencies:

• SVT-AV1-Essential (a build with quarter-step CRF is mandatory with Auto-Boost-Essential v2.0+)
• Python 3.12.x or newer
• Vship (AMD/Nvidia GPU) and/or vs-zip (CPU)
• Vapoursynth and vstools, part of vsjetpack
• FFMS2

Lost users can refer to this Vapoursynth install guide (up until "Installing the JET Packages"). If you are still lost, join the Discord server linked below and ask for help!

Arch users can run this single command to install everything:

yay -S svt-av1-essential-git vapoursynth-plugin-vsjetpack vapoursynth-plugin-vship-cuda-git vapoursynth-plugin-vszip-git ffms2

Radeon users can simply replace vapoursynth-plugin-vship-cuda-git with vapoursynth-plugin-vship-amd-git .

Usage

Auto-Boost-Essential can be considered a helper script, as all you need to do is provide an input video file and it will manage everything for you:

python Auto-Boost-Essential.py -i "my_video_file.mp4"

Even though the above command is sufficient to run the script, one may use additional parameters to tweak the experience, for instance:

Parameter	Usage
--version	Print script version
--verbose	Enable more verbosity [Default: not active]
--input	Video input filepath (original source file)
--temp	The temporary directory for the script to store files in [Default: video input filename]
--no-boosting	Runs the script without boosting (final encode only) [Default: not active]
--resume	Resume the process from the last (un)completed stage [Default: not active]
--stage	Select stage: 1 = fast encode, 2 = calculate metrics, 3 = generate zones, 4 = final encode [Default: all]
--ssimu2	Select the SSIMULACRA2 metric model: vs-zip (CPU), Vship (GPU) or auto [Default: XPSNR metric]
--fast-speed	Fast encode speed (Allowed: medium, fast, faster) [Default: faster]
--final-speed	Final encode speed (Allowed: slower, slow, medium, fast, faster) [Default: slow]
--quality	Base encoder --quality (Allowed: low, medium, high) [Default: medium]
--aggressive	More aggressive boosting [Default: not active]
--unshackle	Less restrictive boosting [Default: not active]
--fast-params	Custom fast encoding parameters
--final-params	Custom final encoding parameters

Yes, the script is even capable of resuming unfinished encodes like Av1an, and can also be run with boosting disabled if all you care is the convenience of a SVT-AV1-Essential wrapper.

Known-issue

The encoding process can hang in the last dozen of frames for certain sources. It can be caused by some specific combination of parameters or by the input file being malformed.
The recommended procedure as of now is: to first backup the ivf file, and to then try to resume with a limited set of parameters. When the issue is caused by the source being malformed, it is possible the encode is complete but the process isn't, in which case simply interrupt the process by hand and manually copy the ivf file wherever you want.
Proper workarounds are actively being researched.

Contribute

The script is open to contributions!
Namely, I'm looking to improve the boosting logic to increase consistency further, with little to no additional performance implications.
Code refactors for clean-up and fixes are also appreciated!

Future

Apart from said boosting logic refactor, the following features are being considered:

• Re-introducing metrics skipping for greater speeds
• Comparing SSIMULACRA2 to XPSNR (because vs-zip's implementation is crazy fast and universal anyway due to running on CPU)
• More robust error handling
• Making the CLI better looking

Benchmarks:

Metrics	Speed

Speed may vary depending on your hardware configuration and source resolution.

The above results are not even best-case scenarios. The selected samples are very complex. More gains are expected on your average clip, granted it contains more than one scene at a minimum!

Join us over at AV1 weeb edition!