Release Release 1.9.8 - Optimization · NevermindNilas/TheAnimeScripter

--input will now support:
- .txt files with a list of videos
- A full string with multiple videos: "G:\TheAnimeScripter\input\input.mp4;G:\TheAnimeScripter\input\input.mp4"
Interpolation will now work prior to upscaling in order to increase the output's quality and to be more VRAM friendly.
TAS now has proper multi subtitle and audio support.
--preset & --preset_list: A new and semi-baked in preset configuration system with reusability in mind.

Usage: --preset {name} will allow for all of the arguments declared within the initial system arguments to be saved inside a JSON file and then reused multiple times.

Note: This lacks an updating/removal system for now, so it's a one-off.

Preset list: Simply prints the available presets with their enabled options.
TAS should now have proper CTRL + C support which extends to a graceful exit of the whole operation.

Upscale TRT should now be up to 10% faster
New coroutines for CUDA detection to improve performance.
Decoupling of STDOUT from the main loop will now result in up to 50% performance gains for decoding. This should benefit all processes.
Overall total model size is now significantly lower.
TensorRT Handler functionality has been improved tremendously and is now significantly more robust and feature-rich than before (this will come in handy in the future).
Rewrote some of the chain loop backend.
Dedup with SSIM-CUDA is now significantly faster.
SSIM CPU is now up to 2x faster ( decode limited )
MSE CPU is now 15% faster.
Upgraded Torch: 2.3.1 -> 2.4.1 for the added performance benefits.
Linux lite builds should be significantly smaller now ( looking for testers & debuggers )
Encoding should now be up to 50% more efficient than before, internal results show dedup ssim speeds with encode going from ~280FPS all the way to ~460-480FPSat 1080P with a 13700k.

Currently TAS is slightly limited by STDOUT and STDIN from FFMPEG. I will look to further improve this down the line.
With the current changes in encoding, I can reasonable argue that CPU and subprocessing are starting to become decent bottlenecks.

With the new changes in Decode/Encode, RAM usage will see a hit, I've seen up to ~500mb more usage in RAM, though that's whilst considering that TAS uses roughly 2.5GB during a reasonable workload.

Enable Preview Window checkbox:
- This will utilize the --preview functionality within TAS to display a live preview of the progress, note this comes with a tax in performance.
New Options for Remove Background:
- TensorRT acceleration has been added to segmentation for further boost in performance.
Pre-Render will be even more flexible:
- The Pre-Render functionality within TAS was and still is fairly limited, but with this release you will have a bit more control over the desired output. You will now be able to choose between two baked in functionalities, Lossless ( the current default ) and Quicktime. For most usecases Quicktime is plenty good and there's no need to choose Lossless but feel free to mess around with it.

Some UI Cleanup of features that are no longer planed to be added in the near future.
Improved some background blur related things.
Since TAS now comes with proper CTRL + C support, you can now exit the process gracefully and not worry about annoying alerts and errors.

Preview is still under active development and I am planning to further improve on it in the near future, but attention will now shift towards the Standalone edition which is about 80-90% of the way done.

FULL: Is for NVIDIA users.
LITE: IS for AMD / Intel users.

Full Changelog: v1.9.7...v1.9.8

Provide feedback