Skip to content
Vapourkit Docs

Changelog

Auto-generated from the Vapourkit desktop repository. Do not hand-edit — update Changelog.md in the desktop repository instead.

2.0.0

Section titled “2.0.0”
  • Filters that build TensorRT engines at runtime no longer look like a frozen app
    • A banner names the engine being built, shows progress when the builder reports it, and explains that this is the first run at that resolution; it clears when the build ends, and on every cancel/crash path
    • Engine builds are kill-safe: the engine is written to a temp file and renamed into place, so force-closing mid-build can no longer leave a truncated engine that gets reused as a cache hit and permanently breaks the filter
    • Builds are recorded in the per-item queue log, and the vs-view launch now evaluates the script first so builds happen under Vapourkit’s UI instead of freezing vs-view’s window
    • Covers vs_temporalfix’s TemporalFix (AI) engine builds too — it builds engines its own way, so its existing build log lines are recognized directly (no progress percentage available, so the banner spins)
    • Third-party filters can opt into the same banner by printing [vk-build] begin/progress/end lines to stderr — see “Runtime engine builds” in the Development docs
  • The RIFE and DPIR filter templates now use real TensorRT when the TensorRT backend is selected, instead of ONNX Runtime CUDA
    • vs-mlrt builds those engines by shelling out to trtexec, which the TensorRT pip wheels don’t ship; Vapourkit now installs a trtexec shim that routes the build through its own TensorRT Python API builder (the same one the model importer uses)
    • The first run at each resolution builds an engine (a few minutes, with the banner above); later runs at that resolution start instantly from the cached engine in data/vsmlrt-models
    • The RIFE model packs now also install the rife_v2 model folder, so templates can select the v2 representation with _implementation=2; existing installs fetch it in the background at startup
  • Python packages are now installed to match your GPU vendor, detected automatically at startup
    • NVIDIA: unchanged — CUDA PyTorch, vsjetpack[full,nvidia], and both vs-mlrt backends (TensorRT + ONNX Runtime)
    • AMD: vsjetpack[full,amd] (HIP/OpenCL/Vulkan plugins) with the DirectML backend, and no multi-GB TensorRT/CUDA stack
    • Intel and unrecognized GPUs: vsjetpack[full,cl,vulkan] with the DirectML backend
    • PyTorch-based filters (vs_deepdeinterlace) get CPU PyTorch on non-NVIDIA GPUs instead of being broken — slower, but they run
  • Installs now clean up packages left over from a different GPU configuration before installing (e.g. TensorRT/cuDNN and CUDA PyTorch when moving to an AMD GPU), including the duplicate ONNX Runtime plugin folder that could otherwise win the autoload race
  • The Plugins modal reports “not installed” when the installed package set targets a different GPU than the one detected, so reinstalling repairs it
    • Existing NVIDIA installs are recognized as-is and are not forced through a reinstall; AMD/Intel users who installed the old CUDA-only package set are prompted to reinstall to get the correct packages
  • Set VAPOURKIT_FORCE_GPU_VENDOR=nvidia|amd|intel|unknown to override GPU detection
  • Fix the RIFE and DPIR filter templates failing with "...models\rife\rife_v4.10.onnx" not found
    • The old zip-based vs-mlrt shipped its model zoo next to the plugin DLLs; the PyPI wheels don’t, so the vsmlrt RIFE/DPIR wrappers had nothing to load
    • The needed packs (RIFE v4.10, DPIR, ~75MB) now download to data/vsmlrt-models during plugin install, and existing installs fetch them automatically in the background at startup; generated scripts point vsmlrt.models_path there so pip reinstalls can’t remove them
    • Other RIFE model versions can be dropped into data/vsmlrt-models/rife manually
  • Rework inference backends into self-contained provider modules (electron/providers/)
    • Each backend (TensorRT, DirectML) owns its script codegen, model-file resolution, pip packages, plugin health checks, and engine building in one place; adding a backend (NCNN, OpenVINO) no longer touches the rest of the codebase
    • The DML/TRT header toggle is now a backend dropdown driven by the provider registry, with the same selection available in Settings
    • Backend choice is now per AI-model filter: every filter defaults to “Auto” (follows the app default) and can override it in the expanded filter card; overridden filters show a badge on the collapsed card
    • Generated scripts now include a vk_backend() helper so custom filters follow the app-selected backend; the bundled RIFE and DPIR templates use it (RIFE previously forced DirectML on every GPU, DPIR needed a hand-edited nvidia_gpu flag). On the TensorRT selection, script filters get real TensorRT through the trtexec shim described above
    • Settings, queue items, and workflow files store a backend id (tensorrt/directml) instead of the useDirectML boolean; existing values migrate automatically on load
  • Migrate the entire install path from manual zip downloads to PyPI
    • VapourSynth (R79), vs-mlrt (16.1), BestSource, and all of pifroggi’s plugins (vs_temporalfix, vs_undistort, vs_colorfix, vs_grain, vs_tiletools) now install via pip
    • Native VapourSynth plugins (akarin, vszip, zsmooth, bestsource, …) arrive automatically as dependencies of vsjetpack[full,nvidia] and the pifroggi packages, using the NVIDIA and JET vs-wheels package indexes
    • vspipe.exe and the core runtime now come from the VapourSynth wheel in Lib\site-packages\vapoursynth; plugins autoload from Lib\site-packages\vapoursynth\plugins
    • vsjetpack is no longer pinned to 1.1.0 (the old vapoursynth==72 ABI pin is obsolete)
    • vsview[full] is installed with the main plugin step instead of a separate pinned install
  • TensorRT engine building now uses the TensorRT Python API instead of trtexec (the TensorRT pip wheels don’t ship trtexec)
    • The Import Model dialog still accepts trtexec-style parameters; unsupported flags are ignored with a warning
  • Portable installs that reuse an existing data folder are migrated in place: the old portable runtime, vs-plugins folder, and bundled script modules that PyPI now provides are cleaned up during setup, and the Python environment is upgraded in place
    • Existing TensorRT engines were built with an older TensorRT and need rebuilding — the existing vs-mlrt version-change prompt handles clearing them
  • NOTE: upgrading a setup install from 0.16.x or older starts fresh (the old installer’s upgrade flow removes the data folder) — export your workflows and filters before upgrading, then re-import them
  • Bundled vs_deepdeinterlace (not yet on PyPI) and the Hybrid scripts continue to install as before
  • Groundwork for Linux support: all platform-specific filenames and the site-packages layout are centralized in electron/constants.ts; the pip install phases are platform-neutral, leaving only the Python bootstrap (and FFmpeg/video-compare downloads) Windows-specific
  • Fix update checker falsely prompting nightly builds to “update” to the stable release they were cut from
    • Nightly version suffixes (e.g. 0.16.1-nightly.2026-05-13) broke the version comparison; nightlies are now only offered stable releases with a strictly newer base version
  • Fix vspipe crashing at startup with v3bdg: unable to acquire api3 VSAPI, abort
    • The bundled fft3dfilter.dll build contains its own API3-bridge guard that aborts the process under VapourSynth R79; it is now removed at install (fft3dfilter is unavailable until an API4 build is sourced)
    • Other API3 plugins load fine through VapourSynth’s compat bridge (with deprecation warnings)
  • Replace API3-only bundled plugins with API4 wheels from PyPI: mvtools, CAS, adaptivegrain, WNNM, KNLMeansCL (nlm-cuda), SCXvid, DCTFilter
  • Fix vs-mlrt ONNX Runtime CUDA support: both vapoursynth-mlrt-ort (CPU/DirectML) and vapoursynth-mlrt-ort-cuda ship a vsort.dll and the CPU-only copy always won the autoload race; the redundant CPU-only folder is now removed post-install (the CUDA build bundles DirectML too)
  • Bundled plugins now extract with skip-existing semantics so they can never overwrite pip-managed plugin files (several share identical filenames)
  • Fix fresh installs on NVIDIA GPUs silently starting in DirectML mode
    • A race persisted useDirectML=true to localStorage before async CUDA detection resolved, permanently blocking the detection-based default
  • Fix DirectML failing with open ..._fp16_fp16.onnx failed when a TensorRT engine model is selected
    • The engine→ONNX path mapping now understands the doubled precision suffix of custom-built engines and picks whichever ONNX candidate exists on disk
  • Pre-included models now get the same ONNX auto-detection as custom imports when opening the build modal
    • Temporal frame count, precision, and static shapes were previously hardcoded (15 channels for any VSR model, precision from filename only), and the frame count was missing from the form entirely

0.16.1

Section titled “0.16.1”
  • Fix Cannot read properties of null (reading 'execute') crash when canceling or restarting an upscale during the frame count probe
    • Same fix applied to the preview-segment path
  • Stream BestSource indexing progress during the frame count probe so cold-cache runs don’t look like a hang
    • Indexing progress now shows in the same progress bar used on first video load, and is written to the queue item log

0.16.0

Section titled “0.16.0”
  • Auto-install plugins at the end of setup
    • Removes the manual “reinstall your plugins” step required by 0.15.0
    • Auto-retries once on transient failure; falls back to Retry / Continue-without-plugins on hard failure
  • Add Privacy mode (lock icon in the header)
    • Hides preview frames, input/output filenames, queue thumbnails, and queue item names behind clickable veils
    • Notification toasts become generic so filenames don’t leak to screen
    • Console auto-collapses when privacy is enabled
    • Setting persists across launches
  • Add descriptive output filenames (enabled by default) — thanks @fs10102020!
    • See 0.15.1 entry below for details
  • Add no-filters safety
    • Persistent banner above the Upscale button when no filters are enabled
    • Confirm dialog before upscaling with zero filters
    • Removed the old “default-upscale” silent fallback that would secretly run whichever AI model was selected first
  • Add BestSource indexing progress bar under the video drop zone on first video load
  • Rename Temporal Fix filters
    • Temporal Fix V2 → TemporalFix (AI)
    • Temporal Fix → TemporalFix (Classic)
  • Fix “Failed to initialize VSScript” on fresh installs
    • Pinned vapoursynth==72 and vsjetpack==1.1.0 so pip doesn’t silently upgrade to an ABI-incompatible Python binding
  • Fix vsview failing to launch (switched from python -m vsview to vsview.exe)
  • Fix descriptive-naming regen ignoring the configured default output folder
  • Fix duplicate video-index-progress terminal event in the get-video-info handler
  • Fix content-length parseInt type error under newer @types/axios

0.15.1

Section titled “0.15.1”
  • Add descriptive output filenames (enabled by default)
    • Output filenames now reflect your workflow instead of using a generic _processed suffix
    • Example: EpisodeName-colorimetry_denoise_4x_resize2160.mkv
    • Includes applied filters, AI model scale, and output resolution
    • Automatically truncates to 32 characters if too long
    • Manually selecting an output path disables auto-generation for that file
    • Toggle available in Settings under Processing
  • Fix TypeScript compilation error in electron/vsMlrtManager.ts

0.14.0

Section titled “0.14.0”

This release in in dedication to my Mom. She passed away on 1/1/26 after a long battle with small cell lung cancer. Rest in peace

  • Adds over 150 new filters, including many from Hybrid!
  • Replaces the filter selection dropdown with a new modal
    • This has a tag system to make finding filters easier
    • It also has a search!
  • Fixes the lag and focus issues present in previous versions of Vapourkit
    • You can now have 20+ filters expanded in your workflow and it will not slow down!
    • The bug that required alt tabbing to fix is no longer present
  • Adds vse-previewer! This allows for realtime previewing of how your video will turn out without having to render the whole thing
    • Replaced vse-previewer with vs-view, a much more modern solution that has more features and is more robust
  • Adds ESC button support to all pop up modals
  • Added vs_grain
  • Replaces pop up dialogs with notifications within the GUI
  • Change legacy “TSPAN” text to “VSR”. This change was made in conjunction with releasing TFDAT, which effectively replaces TSPAN + TSPANv2
  • Lots of GUI tweaks and bug fixes to make it more cohesive and consistent
  • Add option to Settings to set a permanent output path for all videos
  • Add option to duplicate queued items, and overhaul the behavior of the queue button

0.12.2

Section titled “0.12.2”
  • Fix BF16 engine names (previously appended _fp16 when it’s _bf16)
  • Remove unused code
  • Hide Validate button during processing
  • Rename Color Matrix to Colorimetry as it does more than the name implies
  • Improve GUI responsiveness
  • Change the way Developer Log works. It now polls main.log instead of printing directly to the UI
    • This also has the added benefit of fixing formatting issues that were present previously
  • Include 2x_bndl_animefilm_v1.5 FDAT

0.12.1

Section titled “0.12.1”
  • Overhaul validation method. It will no longer automatically run in the background, instead you must manually run it if desired
  • Fix issue where “Same as Input” was the default for fresh installs of Vapourkit
  • Fix broken AV1 presets

0.12.0

Section titled “0.12.0”
  • Remove simple mode to (ironically) simplify codebase
  • Move encoding settings from Settings panel to the right pane, and add easy toggles for common settings
  • Add RIFE filter for frame interpolation
  • Overhaul vkfilter parsing to be more robust
  • Fix GUI design inconsistencies
  • Reverted to vs-mlrt 15.13 as 15.14 has noticeably lower performance
  • Update vs_tiletools
  • Update zsmooth to 0.15

0.11.0

Section titled “0.11.0”
  • Change the way file names are handled for models
  • Overhaul the design of the header to save space
  • Move the DirectML toggle from Settings to the header
  • Change the default model type from vsr to image to reduce chance of error for models without metadata
  • Fix audio clipping when using segments
  • Allow users to customize video-compare settings in the Settings menu
  • Force kill trtexec and vspipe processes when beginning workflow processing
  • Add MC_Degrain filters
  • Change to vs-mlrt version 15.14 from 15.13 RTX
  • Add detection for vs-mlrt version changing (will not take effect in this release)
  • Add BF16 toggle when building TensorRT models
  • Add automatic static + shape detection when building TensorRT models
  • Add update system for vs-mlrt plugin
  • Update vs_undistort to version 2.0.0 (thanks tepete!)
  • Update queue panel behavior and design to be more intuitive

0.10.2

Section titled “0.10.2”
  • Implement segment selection. Users can now select a small segment of a video to process and preview!
    • When using this mode, the comparison buttons are disabled
  • Fix issue where highlighted code wasn’t visible in the Filter panel
  • Minor bug fixes
  • Fix GUI lag
  • Add search function to Manage Models menu

0.10.1

Section titled “0.10.1”
  • Redesign “Show Queue” button and change location
  • Allow the user to change the color space the output video is saved in
  • Rework the Settings menu to be easier to use
  • Fix the way videos are displayed when processing is complete

0.10.0

Section titled “0.10.0”
  • Add batch video processing support
  • Add ability to launch comparisons in from queue list
  • Add experimental update checker
  • Add force stop button for stuck processes
  • Clean up About menu
  • Improve changelog display
  • Fix processing bug with batch processing
  • Update zsmooth plugin to 0.14
  • Overhaul internal code for start/stop processing button
  • Overhaul Video Info Panel
  • Add documentation for Batch Processing
  • Shrink queue panel and clean up unused files
  • Fix color scheme of syntax highlighting

0.9.4

Section titled “0.9.4”
  • Clarify precision options in GUI
  • Add syntax highlighting for filters
  • Add section for license information of included models
  • Add link to GitHub page in About window

0.9.3

Section titled “0.9.3”
  • Fix Logo in header being misaligned in Simple Mode
  • Fix program icon being missing

0.9.2

Section titled “0.9.2”
  • Fix race condition with filters
  • Fix “Start Processing” button not working when Advanced mode AND TensorRT mode are enabled without any built engines

0.9.1

Section titled “0.9.1”
  • Expose previously forced ffmpeg arguments to be edited
  • Remove automatic CUDA detection, turned out to be a driver based issue
  • Add menu to manage models (modify metadata, change precision, rename, delete)
  • Refactored `main.ts

0.9.0

Section titled “0.9.0”
  • Change preview to PNG from mJPEG to improve compatibility and avoid YUV errors
  • ACTUALLY fix —fp32 being added to trt build command
  • Move ffmpeg settings to Settings menu, remove old config file
  • Add automatic detection for CUDA versions, and install different Pytorch versions depending on that

0.8.9

Section titled “0.8.9”
  • Update VapourSynth and filter templates (thanks tepete)

0.8.8:

Section titled “0.8.8:”
  • Add custom engine build command support for tensorrt
  • Rework “Import Model” interface
  • Hopefully fix scrolling bug on right pane when processing a video

0.8.7:

Section titled “0.8.7:”
  • Add labels on header buttons
  • Relabel certain buttons to make their function clearer
  • Prevent processing when ONNX model is selected in TensorRT mode
  • Fix model auto select after building engine

0.8.6:

Section titled “0.8.6:”
  • Fix progress bar in setup screen, round ffmpeg download to nearest integer
  • Fix plugins being missing
  • Fix workflows not notifying the user of missing models
  • Fix workflow names including the extension when loaded

0.8.5:

Section titled “0.8.5:”
  • Static engine support
  • Adds version number to about menu and window title
  • Fixes ffmpeg and vspipe handling when stopping processing, prevents corrupt files
  • Added animations and progress bar text when ffmpeg is stopping
  • Adds MOV as an output option
  • Rolled back to version 0.12 of zsmooth to fix temporalfix
  • Fixed visual bug with num_streams slider