The Anime Scripter - Release [2.0.1]

The Anime Scripter:

New

• Added rtmosr-directml and rtmosr-tensorrt upscaling method by @umzi2.
• Added slow_nvenc_av1 encoding preset.

Improved

• TensorRT and DirectML based depth maps have been significantly improved.
• Changed TAS.log -> TAS-Log.log for easier debugging.
• Changed ffmpeg.log -> FFmpeg-Log.log for easier debugging.
• Generalized some logs.
• Removed unwanted terminal prints generated by ffmpeg when encoding with x265
• Imense improvements in the realtime playback, with the amazing tool from mpv realtime preview of the content is now plausible.
• Switched from FFmpeg to MPV playback for --realtime
• If --realtime is active encoding will no longer take place in order to save on system resources.
• Improved the TRT Shape printing when building engines to be a bit more compact.
• Improved the handling of the ffmpeg encode command.
• Improved reusability of repeated calls, specifically logging and printing.
• Added a forced sys.exit() if FFMPEG crashes during encoding for any reason in order to avoid deadlocks.
• Changed sub from srt -> copy, which should retain metadata from the base input.

Removed

• --depth_quality has been deprecated for directml and tensorrt, now all models will run internally at 518x518 to better match the model resolution and quality.

Fixed

• Shape issues with Segment Anime TensorRT
• Issues with rawvideo and celux
• Issues with 16bit workflow and nvenc_h264

Adobe Edition

New

• Added Take a Screenshot to the Toolbox.
• Added an About Panel which replaces the previous dialog on the top right with more info regarding TAS.
• Added System Requirements in the About panel.

Improved

• Improved the Extracting TAS... indicator, the ... should now 'progress' just so that users know it's not actually stuck.
• Slight adjustment to the backend for ease of reusability, more work needs to be done in order to make TAS slightly more manageable.
• Defaulted depth quality to low
• Added error handling when the desired input was an image, TAS should now generate a toast alert letting the user know that it only works with images.
• More tweaks to how UI elements get generated.
• Improved descriptions of many features available in TAS.
• Significant improvements to the UI/UX, adding labels, icons and descriptions.
• Added direct credits to the Anipy-CLI team in the download an anime section.
• Added direct credits to the ytdlp team in the youtube downloader section.
• Changed the update functionality from a dialog to a toast.
• Refactored even more code.

Fixed

• The Remove Background segment would not update upon selection of different elements.
• Fixed a repeated unncessary task which will now happen only on the initial mount of the UI.

Full Changelog: v2.0.0...v2.0.1

The Anime Scripter - Release [2.0.0]

TheAnimeScripter Version 2.0.0 - 2024-12-29
Discord: https://discord.gg/GVHrTQRMNq

New

• Partially ported TAS to Adobe Premiere and some of the core functionality should now take place.
• Added Full CUDA Graph Acceleration for all TensorRT workflows and some CUDA workflows, this should bring an uplift of up to 25% in performance depending on the input resolution.
• Reverted back to threaded decode / encode due to extremely high CPU usages. This change will imply significantly improved start up times ( up to 1.5s on my system ), significantly less CPU usage ( seen drops of up to 40% on my 13700k ).
• Switch entire decode backend to Celux-CPU ( soon Celux-Cuda )
• Added Rife 4.25 Lite and 4.25 Lite TensorRT
• Added Rife 4.25 Heavy and 4.25 Heavy TensorRT
• Added variable --decode_threads for more granularity.
• Added anime1080fixer-directml support for restoration.
• Added slow_nvenc_h264 & slow_nvenc_h265
• New --restore_method, fastlinedarken and fastlinedarken-tensorrt which is losely inspired by the original Vapoursynth version.
• Added support for dynamic upscale_factor, only works with CUDA workflows.
• Experimental new feature --realtime which will allow real time playback of the output, this is a superset of --preview.
• Thanks to TNTWise, GMFSS has been added back to TAS.
• Added Dynamic Flow Scale, use the command --dynamic_scale, this only works with CUDA for now and is experimental.
• Added temporary metadata collection of the video input.
• Added Flow based deduplication with FlownetS ( experimental )

Improved

• Replaced the existing progress bar with Rich Progress for faster start-up times and more accurate FPS, closer to the ground truth.
• Improved inferrence speed for 16bit specific workloads by up to 50%
• Improved error handling when the totalFrames count was off.
• Improved norm efficiency through removal of repeated operations.
• Upgraded to TensorRT 10.7.0
• Refactored TRT Handler Backend for more granularity and better prints / logging.
• Depth Quality Low will now produce better quality outputs.
• Multithreadded the Encode.
• Simplified some cuda related logs and calls.
• Removed redundant code.
• Changed resizing logic from ffmpeg to Celux ( ffmpeg )
• Improved robustness of --half argument and converted it from boolean to str using the default True; choices [True, False]
• Segmentation ( CUDA ) is now 50% more efficient than before.
• Bump Pytorch 2.5.0 -> 2.5.1
• Error handling for Spandrel.
• Total Frames will now dynamically adjust based on the total amount of decoded items on forced exit.
• Changed logging from Python Based stdout to FFMPEG based stdout.
• Updated Python from 3.12.7 -> 3.12.8
• Added PyMediaInfo fallback if Celux can't detect the metadata.
• Added OpenCV fallback if Celux can't decode.
• Robustness of the I/O handler.
• Added a torch.fx model optimization backend which should improve CUDA inference speeds by up to 20% in certain workloads.
• Upgraded numpy version to 2.2.1

Notes

• Due to the TensorRT Upgrade, all engines will need to be regenerated.

Removed

• Polygraphy in favor of built in TensorRT API.
• --sample_size command.
• pre-compilation of TAS' bytecode.
• removed image support.
• Alive-Progress in favor of Rich-Progress.

Fixed

• Fixed Issues with Rife NCNN.
• Fixed Cuda Race Conditions when interpolating with TRT at higher than 2x factor.
• Fix audio copying to and from .webm inputs.
• Fixed an issue where Depth TensorRT would cause flickering using low depth quality mode.
• Fixed black outputs with Upscale NCNN.
• Fixed an issue where the first frame was black in DirectML workflows.
• Fixed Rife Elexor issues.
• Couple of import issues for --restore
• Fixed an issue where the script would attempt to use an incompatible --encode_method with the .webm container.
• Fixed a bug where if --output was declared and no other processing method was declared tas would only output the temp file.

Adobe Edition

Notes

• This version of TAS is now built against AE 2025, with official support provided. The overall UI and feel have been adjusted to match AE 2025. This may cause mismatches with older versions of AE, and some core functionalities may not work as well or at all.
• The new update checking system does not automatically download TAS Adobe Edition, the end user will still have to manually download it but eventually a fully automated process is planned.
• Toasts do not fully replace alerts, they just serve as a partial replacement to the annoying alerts.

New

• Added an auto Update Cheking system which prompts the user if a new version has been released.
• Added proper Pre-Comp support through internal checks in order to avoid user errors.
• Added new interpolation methods.
• Added new encoding methods.
• Added quick action buttons for Add adjustment layer & Add Solid Layer
• Added Arrange Layers function to quickly sort layers based on user requirements.
• A bunch of tooltips for the end user.
• Added an full path indicators where TAS' backend will be downloaded.
• Added Toasts as an alternative to alerts, this should be less intrusive and less annoying than alerts.
• Added Open Changelogs button.

Improved

• Some naming schemes
• Improved error handling.
• Improved the look and feel of the UI.
• Overall performance and snappiness.
• Upgraded some backend dependencies for extra goodies.
• Improved the preview windows responsiveness.
• Added the new AI options into the dropdowns.

Removed

• Rife 4.18 Cuda and TensorRT.

Fixed

• An issue where TAS would delete files if delete pre-rendered file was turned on but pre-render was not.
• Fixed a bug where there was a 404 image not found error when using preview window

What's Changed

• deps: bump tensorrt from 10.6.0 to 10.7.0 by @dependabot in #62
• deps: bump tensorrt-cu12-libs from 10.6.0 to 10.7.0 by @dependabot in #59
• deps: bump numpy from 1.26.4 to 2.2.1 by @dependabot in #63

New Contributors

• @dependabot made their first contribution in #62

Full Changelog: v1.9.11...TAS_2.0.0

Release [1.9.11]

The Anime Scripter

New

• Added H266 encoding support using command --encode_method h266.
• Added slow_x264, slow_x265 & slow_av1 encode methods.
• Added rife_elexor and rife_elexor-tensorrt
• Renamed --denoise to --restore in order to have a more fitting umbrella term.
• Added Anime1080Fixer by @Zarxrax to --restore
• Added early TRT backend to --restore including anime1080fixer-tensorrt
• Added support for yuv420p10le inputs.
• Added anipy support, use keyword --input "anime"
• Added support for yuv422p8le inputs.

Improvements

• A serious restructure to TAS' backend and I/O handling.
• Upgraded Github Actions runner from Python 3.12.4 -> 3.12.7
• Avoid unnecessary memalocs for decode.
• --upscale_skip will now default to CUDA SSIM or CPU SSIM based on the selected --upscale_method
• --upscale_skip should now be significantly more reliable than before.
• Removed redunant calculations in MSE deduplication improving the performance by ~5%
• Improved RIFE TRT reliability.
• Upgraded Pytorch from 2.4.0 -> 2.5.0
• Improved Input/Output handling which should result in a much more dynamic encoding selection for batch processes.
• Slight adjustment to chained processes should yield in a small performance boost.
• Lazifying TAS has improved start up times by up to 1s in certain workloads, this should improve the responsiveness of TAS by a margin of 30% on average.
• Improved and streamlined TAS benchmarking.
• Upgraded ONNX Opset version 19 -> 20
• Improved Rife Model / Arch Handling.
• Removed unneeded code from decode options.
• Upgraded TensorRT from 10.3.0 -> 10.5.0

Removals:

• Vidgear package
• Video Stabilization.
• TAS Banner, might be reimplemented back in the future.
• Command --hide_banner

Fixed

• 16bit workflow wouldn't work due to undeclared variable.
• Fixed an issue where RIFE TRT would produce unwanted results on the 2nd frame of the output video when CPU usage was high.
• Fixed an issue where RIFE TRT would display duplicated frames past 2x interpolation.
• Fixed benchmark.py issues caused by backend changes.
• Models weights wouldn't be properly loaded for flow caching when using Rife TRT
• Fixed an issue where 16bit workflow would downsample to 8bit then upsample back to 16bit.
• Fixed .mov input issues

Breaking Changes:

• If you are upgrading from an older version of TAS, all engines will need to be rebuilt due to the TensorRT Upgrade.

Notes

• H266 encoding requires FFMPEG version 7.1, if you have FFMPEG in System Path make sure the versions matches, otherwise if you downloaded it using TAS, make sure you delete the ffmpeg folder located at %appdata%\Roaming\TheAnimeScripter\ and allow TAS to re-download the latest build.

Adobe Edition

New

• Added Slow X264 & Slow X265 Encoding options.
• Added Purge Cache button in the Extra Tab which will automatically purge all of AE's cache.
• Added Open TAS Folder button which will open TAS' Backend folder for faster debugging purposes.
• Added Rife Elexor and AnimeFixer to the option list.
• Added Delete Pre-Rendered File Checkbox.
• Added Get Anime Clip button.

Improved

• Restructured some of the code to be more resource efficient, all relevant calls will now only happen once on UI Mount reducing overall CPU usage.
• Some small UI adjustments.
• Improved descriptions.
• Renamed denoise -> restore.
• Pre-Render now dynamically adjusts the cut-off point for the selected layer(s). If the layers extend beyond the bounds of the composition, the part that isn't in the composition will be excluded from processing.

Fixed

• 'Unable to set workareadDuration. Value x out of range y and z.
• Some options were not saved.

Notes

• Temporarely disabled " Remove Background " until some fixes take place.

Release [1.9.10] - Day 0 Patch

The Anime Scripter

Fixed

• Issues where if --inpoint and --outpoint wasn't declared, FFMPEG would self-destruct.
• Issues where sharpening would go out of bounds relative to the max allowed sharpness from FFMPEG.

Release 1.9.9 - Improvements

The Anime Scripter

New

• Enhanced encoding within TAS now allows real-time video processing and viewing with adequate hardware.
• Added Rife 4.25 and Rife 4.25-TensorRT.
• Reintroduced VP9 and QSV_VP9 encoding methods to support .webm inputs.
• More accurate progress information post-process for comparison purposes.
• All executables now include precompiled .pyc code to improve startup times.
• Restructured interpolation pipeline for TensorRT, yielding up to 25% performance gains.
• Full CUDA workflow now uses Cuda Pinned Memory Workflow to increase data transfer efficiency.
• Upgraded Span and Span-TensorRT models to Spanimation-V2 - special thanks to @TNTwise.

Improved

• Rife TensorRT changes significantly reduce the weights folder size. It's recommended to delete the folder at %appdata%\Roaming\TheAnimeScripter\weights and let TAS regenerate all Rife Engines.
• Enhanced metadata retrieval powered by pymediainfo.
• Benchmarking now provides significantly more accurate results.
• Audio & Subtitle Merging now occurs during the encoding phase, reducing overall render time.
• Encoding is now fully parallel to inference, minimizing bottlenecks in the inference thread.
• Improved logging from FFMPEG.
• Enhanced CUDA and encoding handling results in a 30% performance boost for NVIDIA users.
• TAS now attempts to create the folder automatically if --output is specified and the path doesn't exist.
• More robust Try/Except statements reduce code initialization time.
• Improved CUDA synchronization for better reliability.
• Renamed ffmpegLog.txt to ffmpeg.log.
• Renamed log.txt to TAS.log.
• System checks are now more minimal and log less information.
• System checks are now conditional based on the --benchmark flag.
• Slight adjustments to getFFMPEG.py.
• Improved Upscale-Cuda Code.
• Improved Initial loading times by up to 25%

Removed

• Removed --audio and related features in favor of pymediainfo's automated detection.

Fixed

• Corrected exaggerated performance reports from TAS.
• Fixed a regression from 1.9.8 that significantly impacted encoding performance for low end systems.
• Fixed an issue where the first frame of a processed video was black when upscaling.
• Fixed a regression from 1.9.8 where frames were not properly processed.

Adobe Edition

New

• Added Manual Method button for a step-by-step tutorial on manually installing TAS' backend in case of download errors.
• Added Rife4.25 and Rife 4.25-TensorRT Options

Improved

• Enhanced backend download functionality.

Fixed

• Fixed backend download for TAS (Counter: N/A).

Improvements

• Slight code adjustments for better memory efficiency, though benefits are likely minimal.
• UI adjustments to the preview when URL request would 404/500.
• View Border has been defaulted to different style to match the overall UI Theme
• Preview Image should no longer overlap with the border.

Full Changelog: v1.9.8...v1.9.9

Release 1.9.8 - Optimization

The Anime Scripter

NEW

• --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.

Improvements

• 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.

Notes

• 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.

Regressions

• 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.

Removals

• Scikit-Image package.

TAS - Adobe Edition

New

• 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.

Improvements

• 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.

Notes

• 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

Release [1.9.7] - Hotfix

The Anime Scripter

NEW

• Added Rife4.20-NCNN, 4.21-NCNN, 4.22-NCNN, 4.22-Lite-NCNN

Fixes

• CUDA errors seen in the Lite Edition of TAS for Windows.

Regressions:

• Some performance gains found in 1.9.6 have been lost due to errors in dependency handling.

Adobe Edition

New

• Added new NCNN Models

Improvements:

• TAS backend downloads will now be up to 2300% faster.
• Fix some race condition issues found in the backend.
• Made sure for 7z to overwrite existing TAS installations
• Improve path handling
• Better error handling.
• In case of download errors TAS should now attempt to redownload the backend.

Removals

• Removed some old 4.15-lite & 4.17 models in favor of newer and more performant options.

NOTES:

• TAS Full is meant for NVIDIA users who want to benefit from CUDA and TensorRT Accelerations.
• TAS Lite is meant for AMD / Intel users only.

Release [1.9.6] - Linux Edition

The Anime Scripter

NEW

• Added Rife4.22-Lite and Rife4.22-Lite-Tensorrt
• Added --static flag to force static engine generation for Upscale TensorRT.
• Temp state for downloaded models will now ensure file integrity in case of internet loss.
• Added --depth_quality flag with options low, high ( default high ). For the users with lower end system to benefit of higher inference speeds. ( Not reliable for TensorRT )
• Added testing builds for linux-full and linux-lite ( Full is WIP but should be functional with the build-linux.py script )

Improvements

• Upgraded base polygraphy from 0.49.12 -> 0.49.13
• Removed unnecessary clamps.
• The metadata, progress bar and everything else related to video details should be MUCH more accurate now, including logging.
• Log.txt should now be more compact with Arguments no longer having unused arguments.
• Rife TRT will only build Static Engines in order to improve stability and versatility.
• Rife CUDA is now up to 15% faster than before.
• Rife TRT is now up to 15% faster than before.
• Chained Processes with Upscaling and Interpolation TRT / CUDA will now be more memory efficient and performant.
• Benchmarks are now going to be more accurate for interpolation.
• Upgraded ONNXRuntime-DirectML to 1.19.1
• Improved the progress bar with more information and less unnecessary eye candy.
• Improved stability of image inputs.
• More video metadata requests.
• The download progress bar will now also tell the download speed in MB/s
• Some slight adjustments to the building methodology of TAS.

Fixes:

• Some CUDA Race conditions could be met in extremely High FPS workflows ( 500+ FPS ).
• Fix some oversights in Rife 4.22 and Rife 4.22-lite that were reducing the output quality.

Adobe Edition

NEW

• Added Rife4.22-Lite and Rife4.22-Lite-Tensorrt options to the dropdowns.
• Added Depth Quality options ( low and high )

Fixes

• Fixed issues when paths had spaces in them.

Release [1.9.5] - Hotfix Release

This should fix issues with TAS Lite or Full download in mGPU cases.

Release [1.9.4] - Preview

The Anime Scripter

New

• A new 3D like Banner :)
• Added OpenProteus and AniScale2 by Sirosky
• New lite version of TAS.
• Support for multi audio and multi subtitle extraction and addition.
• Added Differential | Differential-TensorRT ( My own attempts ) at a scenechange method ( WIP )
• Added Shift_LPIPS-TensorRT Scenechange Methods by Sudo.
• New --hide_banner flag to Hide TAS' banner when processing.
• Preview functionality with the command --preview, use port 127.0.0.1:5050/frame in order to check the results of TAS.

Improvements

• Improved parameters and examples.
• Upgraded to nightly ONNXRuntime DirectML in order to avoid IR issues.
• Added recursive checks for Model Downloads.
• Significantly better I/O folder handling, this should result in better edge case handling.

Fixes

• Fixed a bug with 16bit workflow and x264_animation_10bit
• Issue #44
• Fixes issues with input folder and output folder cases.
• Issue #46 - Fixed concurrency issues associated with Rife TensorRT and high cpu usages.
• Fixed an issues with span-directml where it would point to a different host.

Regressions

• Temporarely reverted back to pytorch 2.2.2 and cuda 12.1 until .dll issues are fixed, this may result in slight performance losses.

Removals

• Updater.exe

Adobe Edition

New

• TAS will now dynamically switch between "Full" or "Lite" mode depending on your available GPU.

Improved

• Improved Error handling.
• The terminal ( command prompt ) will now start minimized in order to not disrupt your workflow.
• Immense improvements on output handling, TAS will now selectively create and choose specific types of folders based on your current process. This also comes with nice text formatting for the output with keywords like "_DE" for Deduplication and "_SH" for Sharpen.
• Better download functionality with a possibility in increased max download speed.

Bug Fixes

• For some users the download would often hang, this should now have been remedied or at least improved and TAS should now attempt to reconnect in case of a error connection.