Skip to content

Latest commit

 

History

History
680 lines (525 loc) · 20.9 KB

CHANGELOG.MD

File metadata and controls

680 lines (525 loc) · 20.9 KB

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog and follows Semantic Versioning.

[Unreleased] - YYYY-MM-DD

Added

  • Test(s)
  • Slowmotion support through --slowmo

Changed

  • Removed h266

Fixed

  • Fixed more issues with non utf-8 compliant paths. [Thanks Sergio]

Adobe Edition

Added

  • N/A

Changed

  • Small UI redesign for Extra and Toolbox
  • "Path Errors" will now result in toasts rather than annoying alerts.
  • Added slowmo button to the interpolation tab.

Fixed

[2.1.2] - 2025-03-14

Added

  • N/A

Changed

  • More error handlers for easier debugging

Fixed

  • An issue where utf-8 decoding would break Adobe Edition.

[2.1.1] - 2025-03-13

Added

  • 2 New Depth Models, distill_small_v2 and distill_base_v2 which provide better quality with better performance.
  • medium depth quality for better output results for marginal performance loss.

Changed

  • Small tweaks to depth maps which should generate better outputs overall.

Fixed

  • A bug within the download models importing

[2.1.0] - 2025-03-12

Added

  • An automated error handler for when the user wants to use a model which does not work on their system ( WIP ).

Changed

  • Removed unnecessary clamping functionality on decode
  • Updated the mpv configs.
  • Lowered the base strenght of fastlinedarken.
  • psutil and gputil as an temporary alternative to wmi.
  • Upgraded TensorRT to 10.9.0 adding Nvidia Blackwell ( RTX 5000 ) support.
  • Slightly simplified main.py
  • Added more logging for when the user processes a directory rather than a simple file.
  • Slight adjustments to the code to be more resource efficient.
  • Minor improvements to initial launch time ( ~150ms )

Fixed

  • Performance regression when encoding
  • Fixed some padding issues with rife4.25-lite

Adobe Edition

Added

  • Added a progress bar during inference.
  • Added a Logs panel for easier debugging.
  • Added a progress circle when TAS is processing.

Changed

  • Huge chunks of UI generations are now dynamically imported and loaded to make maintaining easier and potentially improve performance.
  • Fully rewrote the Toast generation from the ground up to be more maintainable.
  • The command prompt will no longer show up when processing.
  • Fully redesigned the About Tab.
  • Many minor UI Tweaks.
  • Added an example for segmentation.
  • Changed the run button from only having a logo to an actual text Run Chain which is also slightly separated from the rest of the UI elements.
  • Run Chain will now be disabled during processing.

Fixed

  • N/a

[2.0.5] - 2025-02-17

The Anime Scripter

Added

  • N/A

Changed

  • Bumped numpy version

Fixed

  • Fixed a bug in TensorRT shape parsing when resolution is above 1920x1080

Adobe Edition

Added

  • N/A

Changed

  • Added a fallback to lossless when QuickTime isn’t available

Fixed

  • N/A

[2.0.4] - 2025-02-16

The Anime Scripter

Added

  • Updated rtmosr models to @umzi2’s "v2" models

Changed

  • Bumped Python to 3.12.9
  • Improved compression logic for better results:
    • TAS FULL reduced to 1.7GB (from 1.82GB)
    • TAS LITE reduced to 155MB (from 186MB)
    • Overall extracted filesize decreased by ~100MB
  • Improved subprocess performance in Python 3.12.9 (approx. 0.25s faster)

Removed

  • Removed unneeded files and libraries from TAS

Fixed

  • N/A

Adobe Edition

Added

  • Added lossless encode method
  • Added tip(s)
  • Enforced a Pre-Render step to avoid potential bugs
  • Render Queue panel now auto-closes after rendering
  • Added Add Null Layer button
  • Added a visual indicator for inference speed
  • Added dropdowns in the About Tab for System Requirements and GPU Compatibility

Changed

  • Moved all Pre-Render options to the Encode Settings dialog
  • Simplified backend logic around pre-render
  • Renamed rtmosr to rtmosr v2
  • Added an example to Depth Maps (more options coming)

Fixed

  • Fixed visual anomalies with tooltips

Removed

  • Removed Pre-Render button(s)

[2.0.3] - 2025-02-03

The Anime Scripter

Added

  • N/A

Changed

  • N/A

Fixed

  • Fixed an issue where a static step was still parsed even if the current interpolation method didn’t support it

Adobe Edition

Added

  • N/A

Changed

  • N/A

Fixed

  • N/A

[2.0.2] - 2025-02-02

The Anime Scripter

Added

  • Added ffprobe for more accurate metadata seeking
  • Introduced experimental --static_step command to improve inference speed for Rife CUDA (available for models rife4.6 and rife4.15-lite)

Changed

  • Increased bitrate for slow_nvenc_h265
  • Bumped celux from 0.5.4 to 0.6.0
  • Refactored parts of the ffmpeg code
  • Improved error handling for yt-dlp
  • Added a warning when using anipycli

Fixed

  • Fixed a bug in TRT Handler.py
  • Fixed issue with SVT AV1 generating unneeded login
  • Fixed subtitle issues
  • Fixed padding issues with rife4.25

Removed

  • Removed PyMediaInfo for accuracy reasons
  • Removed the mbps meter from the download progress bar
  • Removed the upscale_skip and interpolate_skip commands

Adobe Edition

Added

  • Added TensorRT Static Engine Generation button in the upscaling settings

Changed

  • Improved error handling
  • Renamed some UI elements

Fixed

  • N/A

Removed

  • Removed Check For updates button

[2.0.1] - 2025-01-12

The Anime Scripter

Added

  • Introduced rtmosr-directml and rtmosr-tensorrt upscaling methods by @umzi
  • Added slow_nvenc_av1 encoding preset

Changed

  • Enhanced depth maps for TensorRT and DirectML
  • Renamed log files (TAS.log → TAS-Log.log; ffmpeg.log → FFmpeg-Log.log)
  • Generalized logging and reduced unnecessary terminal prints from ffmpeg (when using x265)
  • Improved realtime playback with MPV preview and switched playback from FFmpeg to MPV for --realtime
  • Skipped encoding during --realtime to save system resources
  • Enhanced TRT shape printing and improved ffmpeg command handling
  • Added forced exit (sys.exit()) if FFMPEG crashes
  • Changed subtitle handling from srt to copy to retain metadata

Removed

  • Deprecated --depth_quality option (models now run internally at 518x518)

Fixed

  • Fixed shape issues with Segment Anime TensorRT
  • Fixed issues with rawvideo and celux
  • Fixed problems in the 16bit workflow with nvenc_h264

Adobe Edition

Added

  • Added a “Take a Screenshot” feature in the Toolbox
  • Introduced a revamped About Panel with additional TAS info
  • Added System Requirements in the About panel

Changed

  • Improved the “Extracting TAS…” indicator
  • Made minor backend adjustments for better reusability
  • Defaulted depth quality to low
  • Added error handling for image inputs (TAS works only with video)
  • Updated UI elements and descriptions
  • Refactored code for clarity

Fixed

  • Fixed issues with the Remove Background segment not updating correctly
  • Fixed repeated unnecessary tasks on UI mount

[2.0.0] - 2024-12-29

The Anime Scripter

Added

  • Partially ported TAS to Adobe Premiere (core functionality now supported)
  • Added Full CUDA Graph Acceleration for TensorRT and CUDA workflows (up to 25% performance boost)
  • Introduced variable --decode_threads for more granularity
  • Added anime1080fixer-directml support for restoration
  • Added slow_nvenc_h264 & slow_nvenc_h265
  • Introduced new features: --restore_method, fastlinedarken, and fastlinedarken-tensorrt
  • Added support for dynamic upscale_factor (for CUDA workflows)
  • Introduced experimental --realtime mode for real-time output preview
  • Reintroduced GMFSS support
  • Added Dynamic Flow Scale via --dynamic_scale (experimental)
  • Temporarily added video input metadata collection
  • Added Flow-based deduplication with FlownetS (experimental)

Changed

  • Replaced the existing progress bar with Rich Progress for faster startup and improved FPS accuracy
  • Improved inference speed for 16bit workloads (up to 50% faster)
  • Enhanced error handling around ffmpeg calls
  • Upgraded TensorRT to version 10.7.0
  • Refactored the TRT Handler backend for better logging and granularity
  • Optimized normalization by removing redundant operations
  • Simplified CUDA logs and calls
  • Converted the --half argument from boolean to string (default: True; options: [True, False])
  • Improved segmentation efficiency using CUDA
  • Bumped Pytorch from 2.5.0 to 2.5.1
  • Added a torch.fx model optimization backend (up to 20% CUDA inference speed boost)
  • Upgraded numpy to version 2.2.1

Notes

  • Due to the TensorRT upgrade, all engines must be regenerated

Removed

  • Removed Polygraphy in favor of the built-in TensorRT API
  • Removed the --sample_size command and pre-compilation of TAS’ bytecode
  • Removed image support and Alive-Progress (now using Rich-Progress)

Fixed

  • Fixed issues with Rife NCNN
  • Fixed CUDA race conditions at interpolation factors above 2x
  • Fixed audio copying for .webm inputs
  • Fixed depth TensorRT flickering in low quality mode
  • Fixed black output issues with Upscale NCNN
  • Fixed a bug causing the first frame to be black in DirectML workflows
  • Fixed Rife Elexor and related import issues
  • Fixed misuse of --encode_method with .webm containers
  • Fixed a bug where TAS would output only a temp file if --output was declared without other processing

Adobe Edition

Notes

  • This version is built against AE 2025. Some functionalities may not work with older versions.
  • Update checking remains manual for now.

Added

  • Introduced an auto Update Checking system
  • Added proper Pre-Comp support to avoid user errors
  • Added new interpolation and encoding methods
  • Added quick action buttons for “Add Adjustment Layer” and “Add Solid Layer”
  • Introduced an “Arrange Layers” function for easier sorting
  • Added tooltips for enhanced user guidance
  • Added full-path indicators for backend download locations
  • Introduced Toast notifications as an alternative to alerts
  • Added an “Open Changelogs” button

Changed

  • Renamed certain UI elements and improved error handling
  • Enhanced preview window responsiveness and backend reusability

Fixed

  • Fixed Remove Background update issues
  • Fixed repeated tasks on UI mount

[1.9.11] - 2024-10-26

The Anime Scripter

Added

  • Added H266 encoding support via --encode_method h266
  • Added slow_x264, slow_x265, and slow_av1 encoding methods
  • Introduced rife_elexor and rife_elexor-tensorrt
  • Renamed --denoise to --restore
  • Added Anime1080Fixer by @Zarxrax to --restore
  • Added early TRT backend for anime1080fixer-tensorrt
  • Added support for yuv420p10le and yuv422p8le inputs
  • Enabled anipy support using --input "anime"

Changed

  • Restructured TAS’ backend for improved I/O handling
  • Upgraded the GitHub Actions runner (Python 3.12.4 → 3.12.7)
  • Optimized memory allocations and defaults for --upscale_skip
  • Removed redundant calculations in deduplication
  • Upgraded Pytorch from 2.4.0 to 2.5.0 and improved input/output handling

Fixed

  • Fixed variable issues in the 16bit workflow
  • Fixed RIFE TRT duplicate frame issues
  • Fixed benchmark script issues caused by backend changes

Breaking Changes

  • All engines must be rebuilt due to the TensorRT upgrade

Adobe Edition

Fixed

  • Fixed model selection issues between Full and Lite methods

[1.9.10] - 2024-09-29

The Anime Scripter

Fixed

  • Fixed issues where missing --inpoint and --outpoint caused FFMPEG crashes
  • Fixed sharpening values exceeding allowed bounds

[1.9.9] - 2024-09-28

The Anime Scripter

Added

  • Enhanced encoding now supports real-time video processing with adequate hardware
  • Added Rife 4.25 and Rife 4.25-TensorRT
  • Reintroduced VP9 and QSV_VP9 encoding for .webm inputs
  • Precompiled .pyc files now included for faster startup
  • Restructured the interpolation pipeline for TensorRT (up to 25% performance gain)
  • Implemented a CUDA pinned memory workflow to boost data transfer efficiency
  • Upgraded Span models to Spanimation-V2

Changed

  • Reduced the Rife TensorRT weights folder size (recommend deleting the folder to regenerate engines)
  • Enhanced metadata retrieval using pymediainfo
  • Improved benchmarking accuracy and merged audio/subtitle processing during encoding
  • Enabled parallel processing between encoding and inference
  • Improved FFMPEG logging and overall system checks
  • Auto-create output folders if missing

Fixed

  • Fixed exaggerated performance reports
  • Fixed a regression from 1.9.8 impacting encoding performance
  • Fixed the first frame being black in upscaling
  • Fixed frame processing issues from 1.9.8

Adobe Edition

Added

  • Added a Manual Method button for step-by-step TAS backend installation

Changed

  • Enhanced backend download functionality

[1.9.8] - 2024-09-15

The Anime Scripter

Added

  • Extended --input to support:
    • .txt files with video lists
    • Multiple videos in a single string (separated by semicolons)
  • Moved interpolation before upscaling for higher quality and VRAM efficiency
  • Added support for multi-subtitle and multi-audio extraction
  • Introduced --preset and --preset_list for saving and reusing configurations
  • Improved CTRL+C support for graceful termination

Changed

  • Enhanced upscale TRT performance (~10% faster)
  • Implemented new CUDA detection coroutines for better performance
  • Decoupled STDOUT from the main loop (up to 50% performance boost)
  • Reduced total model size and improved TensorRT engine robustness
  • Rewrote backend chained processes and optimized deduplication (both SSIM-CUDA and CPU)
  • Upgraded Torch from 2.3.1 to 2.4.1
  • Improved Linux build sizes and encoding efficiency

Removed

  • Removed the Scikit-Image package

Regressions

  • Noted a RAM usage increase (~500MB) due to new encoding changes

The Anime Scripter - Adobe Edition

Added

  • Enabled Preview Windows using --preview
  • Added new options for Remove Background with TensorRT acceleration
  • Made Pre-Render more flexible with a choice between Lossless (default) and QuickTime

Changed

  • UI cleanup and minor adjustments

Fixed

  • Fixed UI issues and minor scaling problems

[1.9.7] - 2024-09-02

The Anime Scripter

Added

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

Fixed

  • Fixed CUDA errors in the Lite Edition for Windows

Regressions

  • Some performance gains from 1.9.6 were lost due to dependency handling issues

Adobe Edition

Added

  • Added new NCNN Models

Changed

  • Backend downloads now up to 2300% faster
  • Fixed race condition issues and improved path/error handling

Removed

  • Removed older 4.15-lite & 4.17 models in favor of newer, more performant options

[1.9.6] - 2024-09-02

The Anime Scripter

Added

  • Introduced Rife4.22-Lite and Rife4.22-Lite-Tensorrt
  • Added --static flag to force static engine generation for Upscale TensorRT
  • Added temporary state for ensuring file integrity during model downloads
  • Introduced --depth_quality flag (options: low, high; default is high)
  • Provided testing builds for linux-full and linux-lite

Changed

  • Upgraded base polygraphy from 0.49.12 to 0.49.13
  • Removed unnecessary clamps and optimized metadata/progress reporting
  • Made log.txt more compact by removing unused arguments
  • Ensured Rife TRT builds static engines for improved stability
  • Improved Rife CUDA and TRT performance (up to 15% faster)
  • Enhanced chained processes for better memory efficiency and benchmarking
  • Upgraded ONNXRuntime-DirectML to 1.19.1 and improved download progress reporting

Fixed

  • Fixed CUDA race conditions in extremely high FPS workflows
  • Fixed issues in Rife 4.22 and Rife 4.22-lite that reduced output quality

Adobe Edition

Added

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

Fixed

  • Fixed path issues when spaces are present

[1.9.5] - 2024-08-17

The Anime Scripter

  • No significant changes

Adobe Edition

Fixed

  • Fixed issues with Full vs. Lite method selection during model downloads

[1.9.4] - 2024-08-17

The Anime Scripter

Added

  • Introduced a new 3D-like Banner
  • Added OpenProteus and AniScale2 by Sirosky
  • Launched a new lite version of TAS (WIP)
  • Added support for multi-audio and multi-subtitle extraction
  • Introduced ShuffleSpan by Sudo (WIP)
  • Added Differential and Differential-TensorRT scene change methods (WIP)
  • Added Shift_LPIPS-TensorRT scene change methods by Sudo
  • Added --hide_banner flag to suppress the TAS banner
  • Introduced preview functionality via --preview (access via 127.0.0.1:5050/frame)

Changed

  • Improved parameter examples and upgraded to nightly ONNXRuntime DirectML
  • Added recursive checks for model downloads
  • Enhanced I/O folder handling for better edge-case management

Fixed

  • Fixed a bug with 16bit workflows and x264_animation_10bit
  • Resolved issues with input/output folder case sensitivity
  • Fixed concurrency issues with Rife TensorRT and high CPU usage
  • Fixed span-directml host misdirection

Regressions

  • Temporarily reverted to pytorch 2.2.2 and cuda 12.1 due to .dll issues

Removed

  • Removed Updater.exe

Adobe Edition

Added

  • TAS now dynamically switches between “Full” and “Lite” modes based on GPU availability

Changed

  • Improved error handling and minimized terminal disruption (now starts minimized)
  • Enhanced output formatting with keyword highlighting (e.g., _DE for deduplication)
  • Improved download functionality with potential for higher speeds

[1.9.3] - 2024-08-09

The Anime Scripter

Added

  • Added Rife4.22 TensorRT
  • Introduced Video Stabilization using Vidgear

Changed

  • N/A

Removed

  • N/A

Adobe Edition

Added

  • TBA

Changed

  • TBA

[1.9.2] - 2024-08-09

The Anime Scripter

Added

  • Upgraded to Python 3.12
  • Added Rife 4.22 CUDA, Rife 4.21 CUDA and TensorRT, and Rife 4.20 TensorRT
  • Introduced --bit_depth option (choices: 8bit, 16bit)
  • Added experimental .GIF input support via --encode_method gif
  • Added experimental .png/.jpeg support via --encode_method image
  • Overhauled model downloading (recommend deleting the weights folder)
  • Upgraded SPAN models to 2x_Spanimation V1.5

Changed

  • Achieved up to 25% faster Rife TensorRT inference
  • Improved TensorRT engine robustness and reduced available model count
  • Streamlined deduplication and encoding performance

Removed

  • Removed Rife4.15, Rife4.15-ncnn, and Rife4.15-TensorRT

Fixed

  • Fixed TensorRT artifacting in Depth Maps
  • Resolved inconsistent Depth Map outputs and aliasing issues
  • Fixed issues with --custom_encoder
  • Partially addressed issue #44
  • Fixed pathing issues for Linux

The Anime Scripter - Adobe Edition

Added

  • Enabled auto clip functionality
  • Introduced half-automated TAS updates
  • Added --bit_depth to encoding settings

Changed

  • Various UI improvements

Fixed

  • Fixed scaling issues and random TAS update popups

[1.9.1] - 2024-07-30

The Anime Scripter

Added

  • Added Rife 4.20 CUDA
  • Introduced Depth Anything V2 – DirectML
  • Added --scenechange_method (options: maxvit-directml & maxvit-tensorrt)
  • Added ProRes 444 HQ Encoder via --encode_method prores and ProRes encoding support

Changed

  • Improved upscaling performance with TensorRT (up to 15% faster)
  • Improved upscaling performance with CUDA (up to 20% faster)
  • Enhanced interpolation performance with TensorRT (up to 10% faster) and CUDA (up to 30% faster)
  • Decoupled scene change from interpolation with a performance boost (up to 20% faster with DirectML/TensorRT)
  • Optimized Rife UHD mode and overall code formatting
  • Upgraded Torch from 2.2.2 to 2.4.0 and CUDA from 12.1 to 12.5
  • Reformatted help output for clarity

Fixed

  • Fixed issues #43 and #41
  • Fixed scene change caching and benchmark script issues
  • Resolved missing DLL issues with onnxruntime-directml

Adobe Edition

Added

  • Introduced a new secure installation method with code signing
  • Enabled dynamic downloading and updating of TAS
  • Enhanced Pre-Render to support multiple active/selected layers

Changed

  • Redesigned the UI for a more compact layout (WIP)
  • Moved the About page to a Help icon
  • Improved tooltips and progress indicators
  • Reduced the initial executable size (230kb) with Full and Lite versions available
  • Updated dropdown hints and model sectioning
  • Added new models and encoders; changed preset template from lossless to high quality

Fixed

  • Fixed inpoint/outpoint boundary issues
  • Fixed YouTube import issues when no composition was selected
  • Fixed Pre-Render work area adjustment issues