Skip to content

About

Automated AV1 video transcoding pipeline powered by Auto-Boost-Essential and Opus audio encoding, with a real-time web dashboard.

Resources

Stars

23 stars

Watchers

2 watching

Forks

Repository files navigation

Rabbit Encoder

Automated AV1 video transcoding pipeline powered by Auto-Boost-Essential and Opus audio encoding, with a real-time web dashboard.

Drop media files into the input folder and get optimally encoded MKV files in the output folder, or browse your media library directly from the dashboard and encode entire series in-place.

Features

  • Auto-Boost-Essential integration for optimal per-scene CRF zones
  • Opus audio encoding with configurable per-channel bitrates
  • HDR10 metadata preservation (PQ, BT.2020, mastering display, content light)
  • HDR10+ metadata measured for every PQ source with retone and written into the AV1 encode, with the scenes of a Dolby Vision source's shots
  • HDR Metadata tool to inspect, add, measure, import, change and remove HDR10+, HDR10 and Dolby Vision metadata in existing AV1 MKVs, one file or a whole folder, without re-encoding
  • Film Grain tool to inspect, add, replace and remove AV1 film grain synthesis in existing AV1 MKVs, one file or a whole folder, without re-encoding
  • VapourSynth filter chain stackable per-job filters (FineDehalo, DehaloAlpha...) with light / medium / heavy presets, full per-parameter overrides, and a hot-reloadable user preset directory for dropping in your own .vpy scripts
  • Web dashboard for monitoring progress and configuring per-file settings
  • File watcher auto-detects new files in the input directory, skipping the staging files and *.repaired.mkv, *.audio.mkv and *.subtitles-edited.mkv copies Rabbit's tools write
  • Queue system processes files sequentially, prioritizes quick repair jobs over pending encodes, and supports reordering plus pause/resume
  • Preview encoding generates configurable short samples spread across the source, with frame-aligned source/encode comparisons before committing the full job
  • Library encoding browse mounted media folders from the UI and encode in-place, replacing source files
  • Subtitle repair/remux replace, remove, reorder, retag, restyle, or recompress MKV subtitle tracks without re-encoding video or audio
  • Subtitle editor browse MKVs, navigate and edit SRT/ASS/SSA cues, and preview timing changes in an audible video loop (three seconds before and after each cue by default). Adjust track offsets and loop boundaries. Jump to a cue by time or text, edit the whole track in a raw file view, and turn the video preview off for fast browsing. Bitmap PGS/VobSub/DVB subtitles use track delays. Replace the MKV after verification or save a separate *.subtitles-edited.mkv, with video and audio copied unchanged.
  • Font check find the fonts a folder's subtitles use but its videos do not carry, then attach them from an upload or the font library, or point the subtitles at the closest font the videos already have
  • Subtitles from files add a folder of ASS, SSA or SRT files to their episodes, with the fonts they use, uploading and renaming the ones neither the videos nor the folder have
  • Subtitles from releases copy the subtitle languages your encodes lack from other releases of them, each track retimed to the encode by comparing their audio. Works on a folder of encodes or single encodes, with release folders or single release files. Track types lets you say what a track really is, such as Alt Subtitles that are Honorifics, leave a track out, or change its language tag, for every episode at once. Episodes are paired with their releases the same way as in Audio from releases, and every pairing can be fixed by hand
  • Subtitle search finds subtitles for a video on OpenSubtitles.com and SubDL and downloads the most promising ones per language. Each one is synced by rewriting its timestamps to match the video's own subtitles, the display times of its picture subtitles, or the speech in its audio. Offsets, PAL speedups and cut scenes are handled. Every download is rated on how well it fits and whether parts are missing. The best one can be added as a track or saved next to the video. Several episodes can be searched at once. One source, such as CR, can then be picked for every episode that has it, its credit renamed for all of them, and languages missing from any episode left out. Downloads use the OpenSubtitles.com account you log in with under Settings and count against its daily quota. A SubDL API key, free from your SubDL account, adds SubDL to every search, or replaces OpenSubtitles when you have no login. SubDL allows 300 downloads a day. Downloads are cached, so a file never uses the daily quota twice. With Subtitle translation on, missing target languages are downloaded this way first, and only the languages without a good subtitle are translated with AI. Up to 30 of the most used languages can be added on top, one download each, kept only when they sync well. Missing languages can also be added without encoding, from the Library or from Subtitles, Translate subtitles for any video or folder in the input, output or library folders.
  • Picture subtitle OCR reads PGS subtitles with a vision model you choose, such as Surya OCR 2 or Chandra OCR 2 served with vLLM, a model in Ollama, or a cloud model, and adds them as text tracks with their exact timing and italics. Every picture is read once, however often it is shown. The text tracks can be styled like other subtitles and translated, so BluRay files with only picture subtitles can be translated too. Runs as part of encodes and translate jobs, or on its own from Subtitles, Subtitle OCR for one MKV or a whole folder, optionally removing the picture tracks it read. When an encode no longer has its picture subtitles, they can be read from its original, one file or a folder of encodes matched to a folder of originals, and their timing is checked against the encode and shifted if it was cut or delayed.
  • Audio from releases copy the audio languages your videos lack from other releases of them, synced by the audio they share. Offsets, PAL speedups and cut scenes are handled. Added tracks can be encoded to Opus and every audio track ordered, named and flagged as an encode would do it
  • Audio tracks give a folder of existing videos the audio an encode with your settings would have, without touching the video. Tracks are encoded to Opus, removed, reordered and renamed, with a choice per track type
  • Folder metadata audit group MKVs by their ordered audio/subtitle metadata, edit and reorder audio and subtitle tracks across an entire group, highlight missing or inconsistent tracks, and open outliers against the majority layout in the Rabbit repair editor
  • Blu-ray discs split a set of ripped Blu-ray discs (folders with BDMV inside) into episodes, specials and extras numbered as Sonarr does with TVDB, or into a movie and its extras as Radarr does with TMDB, and encode them into the output folder or straight into one of your library folders, such as /Animes or /Movies, with names from your own naming formats (see Disc Naming). The folder is chosen under Save to and remembered separately for series and movies. Whether a set holds a series or a movie is guessed from its playlists and the runtimes TVDB and TMDB give, and can be switched. For a movie, the feature is the playlist closest to its runtime. Where a studio added shuffled copies of the feature, the one that plays the clips in order is taken, and other cuts are offered to keep as editions. Episodes are found from the disc's playlists: one playlist per episode, one play item per episode, or the chapters of one long clip. Menu loops and duplicate playlists are left out, and extras under a minute, mostly copyright warnings and logos, start unticked. Short clips between episodes, such as dub credits, stay with the episode they follow. Series, episode titles and runtimes come from SkyHook, the TVDB service Sonarr uses, and movies from Radarr's TMDB service, so only the title is sent and no API key is needed. Every episode can be checked, renumbered, renamed or left out before queueing, and an episode found at chapters can be moved to start at another chapter. Their encodes are tagged as Bluray in the file name and carry SOURCE BDMV in their metadata, along with a BDMV tag that records the disc, playlist and range they came from. Later encodes of the file keep it. Each item is remuxed from the disc with mkvmerge right before its encode, so the temp folder only ever holds one episode, and the discs are never changed. With Result set to Remux only, each item is written as a lossless MKV under the name in the Output column instead, with the same tags and no encode. When a denoised file made from a disc is encoded on another system, the disc set can be chosen as its original for film grain wherever an original is asked for: in the job's Film grain settings and in Film grain under Blu-ray discs..., and as the originals folder when queueing videos or measuring a folder. Each video is matched to its range by its BDMV tag, or by its episode number and length when it has none, and the range is remuxed into the temp folder only while the grain is measured. Compare takes a disc set too, under Add Blu-ray disc..., matched to the first video the same way or chosen from the set's items. Its remux stays in the comparison's folder for adding frames and measuring, and only the newest comparison keeps it. Disc sets copied into the input folder are offered for review once their files stop growing. Others can be opened from Tools, Blu-ray discs
  • Jellyfin / Sonarr integration automatically cleans up .nfo and thumbnail files when replacing sources so metadata is regenerated
  • Smart skip already-encoded files (detected by -{ORGANIZATION} suffix) are recognized and skipped

Quick Start

# 1. Configure settings in docker-compose.yml (or use defaults)

# 2. deploy container
docker compose up -d

# 3. Open the dashboard (http://localhost:3000)

# 4. Drop files into the input folder
cp movie.mkv input/

Configuration

Settings are configurable via environment variables in docker-compose.yml:

Variable Default Description
PORT 3000 Web dashboard port
PASSWORD rabbitencoder Password to access web dashboard
FILE_COOLDOWN 30 Seconds the file size must stay unchanged before encoding starts
ORGANIZATION RabbitCompany Tag appended to encoded filenames (e.g. -RabbitCompany)
INPUT_DIR /data/input Watched input directory
OUTPUT_DIR /data/output Encoded output directory
TEMP_DIR /data/temp Temporary encode and preview files
LIBRARY_DIRS empty Comma-separated mounted library roots
RECYCLE_BIN_DIR empty Keep originals of in-place library encodes here (see below)
FONTS_STOCK_DIR /app/fonts Seed font groups copied into the user dir on first start
FONTS_USER_DIR /config/fonts User font groups (read/write, the only directory scanned)
SYSTEM_FONTS_DIRS /system-fonts Read-only host font dirs, browsable to import fonts into groups
VS_PRESETS_STOCK_DIR /app/vapoursynth/presets Built-in VapourSynth preset directory
VS_PRESETS_USER_DIR /config/vapoursynth/presets User VapourSynth preset directory
VS_RABBIT_MODULE_DIR /app/vapoursynth Rabbit Encoder VapourSynth helper-module directory
OPENSUBTITLES_API_KEY built in Your own OpenSubtitles.com consumer key instead of Rabbit's
SERIES_FOLDER_FORMAT {Series TitleYear} Series folder of Blu-ray disc imports (see Disc Naming)
SEASON_FOLDER_FORMAT Season {season:00} Season folder of Blu-ray disc imports
SPECIALS_FOLDER_FORMAT Specials Folder of season 0 episodes from Blu-ray discs
EPISODE_FORMAT see Disc Naming Episode file name of Blu-ray disc imports
MOVIE_FOLDER_FORMAT {Movie Title} ({Release Year}) Movie folder of Blu-ray disc imports
MOVIE_FORMAT see Disc Naming Movie file name of Blu-ray disc imports

If ORGANIZATION is set to RabbitCompany (the default), then any file ending with -RabbitCompany.mkv is treated as already encoded and will be:

  • Shown with a green encoded badge and dimmed in the library browser
  • Completely skipped when you click Encode Folder

This means you can safely run Encode Folder on the same series multiple times (only new or unencoded files will be queued).

The Library browser skips recursive folder counts by default, so opening folders on network storage stays fast. It still shows sizes and encoded badges for video files in the open folder. Turn on Show folder counts and encoded status in the Library window to scan subfolders and show video totals and green badges for fully encoded folders. This choice is saved in your browser. Encoding still skips files that are already encoded when counts are off.

If RECYCLE_BIN_DIR is set, the original of every library file that is replaced in place is moved into that directory instead of being overwritten or deleted, keeping its full path:

Library:      /Animes/example/series/Episode 1.mkv
Recycle bin:  <RECYCLE_BIN_DIR>/Animes/example/series/Episode 1.mkv

The recycle bin is never emptied automatically, so delete the originals yourself once you have checked the encodes. Mount it on the same filesystem as your library to make the move instant as otherwise the original is copied. Keep it outside LIBRARY_DIRS.

Encoding Pipeline

For each file, the engine runs:

  1. Download Source: Copy an NFS or SMB source into the local temporary folder before reading its metadata or tracks. The progress bar shows bytes copied and transfer speed in Mbps or Gbps. Local sources are used directly. The copy is skipped when the temporary folder is on network storage or lacks space for the source and a 512 MiB reserve.
  2. Probe: Extract media information such as resolution, audio layout, frame rate, subtitle tracks, and HDR metadata from the working source.
  3. Read Source: Read video timestamps and copy the video stream into the temporary working folder. The detail text shows the progress of each operation.
  4. Prepare: Correct frame timing, run the configured VapourSynth passes, and apply crop, scaling, debanding, and denoising.
  5. HDR10+: For PQ sources, measure HDR10+ metadata of the prepared video with retone.
  6. Auto-Boost-Essential: Run scene analysis, quality metrics, CRF-zone generation, and the final AV1 encode, then write the HDR10+ metadata and film grain into it.
  7. Audio: Sort, filter, deduplicate, and encode selected audio tracks to Opus through a FLAC pipe, or copy them when configured.
  8. Subtitles and fonts
    • Sort, filter, deduplicate, and rename subtitle tracks.
    • Optionally convert SRT subtitles to ASS.
    • Restyle dialogue-classified ASS styles without changing signs, songs, or other untouched styles.
    • Resolve language/script-specific font faces.
    • Instance selected variable-font axes into static font files.
    • Remove unused source font attachments when enabled.
    • Preserve source fonts still referenced by surviving untouched styles or inline \fn overrides.
    • Use numbered aliases such as Noto Sans 2 only when a retained source font already occupies the requested family name.
  9. Mux: Merge video, audio, subtitles, chapters, fonts, and metadata into the final MKV, then apply color metadata. The progress bar follows mkvmerge.
  10. Write Output: Move the finished MKV to its destination. When the destination is on another filesystem, the progress bar shows bytes transferred and speed in Mbps or Gbps, including a source copy to the recycle bin when needed.

Live View

The Live View button on an encoding job opens a view of everything the job has done and is doing: every step with its timing and result, the elapsed time, and the estimated or final size.

With Rabbit Boost, it also shows the bitrate of the fast pass and the final pass second by second, read from the encoder's output while it is written, and the CRF and SSIMULACRA2 score Rabbit Boost planned for each scene. The bitrate expected for each scene is known before the final pass starts, so a scene that would go far over what you want shows up early and the job can be cancelled and its settings fixed. The video size is estimated at every stage: roughly from the fast pass at the base CRF, then from each scene's predicted bitrate once the zones are planned, and from what the final pass has written while it runs. With a maximum bitrate set, the view warns when a scene is expected to go over it or when 5 seconds of the final pass did. Auto-Boost-Essential runs as its own program, so its jobs show only their steps.

With Follow queue, which is on when the view is opened on an encoding job, it switches to the next job once the shown one is over, so it can stay open while a whole season encodes.

Encode Report and Bitrate Analysis

The chart button on a finished job opens its Encode Report. It charts the bitrate of the finished file, colored by the auto denoise level each part got. For a Rabbit Boost encode it adds the max bitrate and a second chart of every scene's CRF and fast pass SSIMULACRA2 score, with the scenes the bitrate cap held back marked. While the source is still there, Measure scores the encode against it over the whole video with SSIMULACRA2, XPSNR and VMAF, with the crop the encode got, and adds a chart of the score with its mean, 5% low and worst moment. Every chart follows the same zoom. On a queued job the same button shows the source bitrate and the auto denoise thresholds instead.

Export data saves everything shown, measurement included, as a .bitrate-analysis.json. Tools > Bitrate Analysis opens the same view without a job: choose any video to chart its bitrate, or import an exported file to see its charts again after the job is gone.

Preview Encoding

Preview encoding creates configurable samples distributed across the source so source and encoded results can be compared before running the full job.

Each sample starts from an exact, lossless FFV1 video window rather than a stream-copied GOP fragment. This prevents keyframe preroll and timestamp offsets from causing source and encoded comparison images to land on different frames. The default is 6 samples, and both the sample count and duration can be changed through the dashboard or preview API.

Intermediate VapourSynth and prepare-filter stills are also exposed when those stages are active.

Film grain tables

Rabbit measures film grain with regrain, which compares the grainy original with the denoised video and writes an AV1 film grain table with an entry for every scene. AV1 draws a new grain pattern for every frame, while part of a source's grain often stays in place from frame to frame. regrain compares the grain of consecutive frames where the picture stands still and scales each scene's grain to match, so it shimmers no more than the source's.

When adding externally denoised videos from the Library, choose an original file or a folder of originals in Queue Videos. Rabbit proposes unique episode or filename matches. Review the pairs and adjust any unmatched rows. Generate FGS tables automatically for mapped videos is on by default when a match is found and can be turned off. Mapped originals are also preselected in Preview. Unmapped videos encode without automatic FGS. Each mapped video measures its table as a Film Grain step of its own encode, right after Analyze. regrain uses every CPU core, so the step never runs beside another job's denoise or encode. The step shows frame progress, speed, and ETA. If generation fails, the job fails with the error, and retrying the job retries generation. Starting a Preview for a mapped job measures the table first when it does not exist yet.

Measurement range controls which part of the video is measured. Quick is the default for new jobs: regrain measures a 30-second window near the middle of both videos, and Rabbit keeps its middle entry for the whole encode. Videos shorter than a minute use the full video. Full video measures the whole video, so every scene keeps its own grain. The resulting table is stored on the job and reused for every Preview and the final encode.

Scan controls how much of the measured range regrain reads. Every frame is the most accurate. Half, Quarter and Eighth read that share of every 48 frames and are faster, with a small loss of accuracy. On an 8 core laptop CPU, a full video measurement of a 24 minute episode took about 4 minutes with Every frame and under 2 minutes with Quarter. Both controls are available in Queue Videos and Advanced Settings -> AV1 film grain table. Changing them does not regenerate a table that already exists. Generate it again to use the new choices.

For an individual queued video, Advanced Settings -> AV1 film grain table can generate a table from its aligned grainy original, or accept an existing table path inside the encoder container. Generation runs right away and shows frame progress, speed, and ETA in that window. It can be cancelled, and it continues if the window is closed. Run it while no encode is active, because it uses every CPU core. The table is stored in the mounted temp directory. The original and denoised files must have the same constant frame rate and start on the same frame. regrain checks both and stops with an error when the videos have a variable frame rate or are off by a frame. When the denoised file was cropped first, Rabbit finds where its frame sits inside the original and lets regrain measure only that area. When the denoiser removed no measurable grain, the table is empty and the video is encoded without film grain. Run Preview before the full encode.

When Rabbit denoises the video itself, turn on Advanced Settings -> AV1 film grain table -> Restore denoised grain. No original has to be chosen. Rabbit keeps the video as it is right before the denoise pass and measures the grain the denoiser removed as a Film Grain step right after Prepare. The kept copy is a hard link, so it takes no extra space. Crop is passed to regrain as the measured area. With Downscale on, Rabbit first writes a scaled copy of that video, so both sides are measured at the encode resolution. Measurement range and Scan apply as above. The table is measured on every run and never stored on the job, so a changed denoise setting always gets a fresh table. A Preview measures each sample clip on its own. When auto denoise finds nothing to denoise, the step is skipped. A table from an original or a table path takes precedence over this setting.

Amount -> Just enough to hide banding puts back only as much of the measured grain as it takes to hide banding, at most a deviation of 0.5 on the 0 to 255 scale. It uses regrain diff --just-enough and applies to every measured table, both from an original and from Rabbit's own denoise.

Pattern -> Static grain pattern applies the table with regrain apply --static-grain. Every frame of a table entry uses the entry's seed, so the decoder lays the same grain pattern over the whole scene like a texture on the screen, instead of a new pattern each frame. It suits light grain on footage that mostly stands still, such as grain that only hides banding in anime. Where the camera pans, the pattern stays in place while the picture moves under it. The pattern changes where the next table entry starts. It applies to every table, also a synthetic one or a table path you enter yourself.

Synthetic grain writes grain of a chosen strength with regrain generate for jobs where no measured table applies, including jobs without any denoise. Strength is given as an ISO or as a Deviation, and 0 turns it off. ISO is the camera sensor noise of regrain generate --iso, the photon noise model of libaom that aomenc and SVT-AV1 use for --photon-noise, so the same ISO gives the same grain. Rabbit gives regrain the frame size of the encode, after the downscale to 1080p, and the whole 16:9 frame when black bars were cut, so a letterboxed film gets the grain of its frame. A smaller frame has larger pixels and less noise: at 1080p, ISO 200 to 600 gives a deviation of 0.38 to 0.50, about what hides banding in anime, and 4K needs a quarter of the ISO for the same grain. ISO grain is weaker in shadows and stronger in highlights, as a sensor's is. Deviation is the luma deviation on the 0 to 255 scale, 0.6 times as strong at black and 0.7 times at white as in the middle, as film grain is. Settings saved before ISO keep their deviation. Fine is single pixels like digital camera noise, Medium is like most film grain, and Coarse forms clumps a few pixels wide. Chroma adds color grain. With Restore denoised grain also on, synthetic grain is used when nothing was denoised or the denoiser removed no measurable grain.

After the video is encoded, Rabbit writes the table into the AV1 stream with regrain apply, for previews and full encodes. regrain then reads the result back, checks every frame's grain, and decodes it with dav1d before Rabbit continues. SVT-AV1's own --fgs-table reads only the first entry of a table, so it is not used. A Full video table is applied scene by scene, and each segment keeps the grain measured for it. A Quick table is applied as one segment to the whole video. A table path you enter yourself is applied as written, so its timestamps have to match the video. Keep the source table available until encoding finishes. Do not combine the table with --photon-noise, --film-grain, or --fgs-table in custom encoder parameters. Grain the encoder wrote itself is replaced by the table.

The container downloads the checksum-verified regrain v0.5.0 release. It is a static binary that uses the container's FFmpeg.

HDR10+ metadata

AV1 encoders cannot measure HDR10+ metadata, so a PQ source used to come out as plain HDR10 even when it had HDR10+ or Dolby Vision. With HDR -> Write HDR10+ metadata, which is on by default, Rabbit measures the metadata with retone as an HDR10+ step after Prepare and writes it into the AV1 encode. HDR10+ displays then tone map every scene on its own instead of the whole video at once. The setting only acts on PQ sources encoded to AV1. HLG and SDR sources are encoded as before.

The step measures the video as it will be encoded, after crop, downscale and denoise, so the metadata describes the picture the display gets. Every scene gets one set of values, so a display does not change its tone mapping in the middle of a shot. When the source has Dolby Vision, the scenes follow the shots its colourist marked, read from the Dolby Vision metadata with retone convert. When that metadata does not fit the prepared video frame for frame, for example after dropped frames were filled, the scenes are measured instead and the log says so. A Preview measures each sample clip on its own.

After the final encode, retone apply writes one HDR10+ metadata OBU right before every shown frame, as the HDR10+ AV1 specification requires, and tags the video as BT.2020 PQ when the encoder left the tags out. When a film grain table follows, regrain's verification decodes the stream with the HDR10+ metadata in it. Otherwise retone reads the result back and decodes it with dav1d itself. When the source states no MaxCLL and MaxFALL, the values the measurement found are written into the MKV colour elements. The step shows the number of scenes and where they came from.

HDR10+ metadata is not worth failing an encode over. When measuring or writing it fails, the step says why, the log has a warning, and the encode finishes as HDR10. The settings code records hp=0 when the setting is off.

The container downloads the checksum-verified retone v0.1.3 release. It is a static binary that uses the container's FFmpeg.

Subtitle Repair / Remux

Open Repair in the dashboard, select an encoded MKV and optionally its original source MKV, then inspect their subtitle tracks. The output plan can keep or remove target tracks, import source tracks, reorder them, edit names/languages and Matroska flags, and choose zlib, uncompressed, or the track's current compression. Text tracks can either be copied unchanged or processed with the current Rabbit ASS/SRT styling defaults.

Repair jobs use MKVToolNix stream copying: video and audio codecs are verified before the result is accepted, and source/target duration mismatches are rejected. The single-file repair editor's Save changes to defaults to Existing MKV file. Choose New .repaired.mkv copy to write a non-colliding copy beside the encoded target instead. Replacement only happens atomically after the staged MKV passes verification.

Use Audit folder to inspect every MKV directly inside a selected folder. Files are grouped as A, B, C, and so on by their ordered audio/subtitle track type, language, title, and flags. Group A is the majority layout. Outliers list their exact differences and open directly in the repair editor. Audits launched from a completed queue folder also pair each encode with its own original input, making missing source subtitles available for plain copying or Rabbit ASS/SRT processing without ever borrowing subtitle content from another episode.

Choose Edit group beside a group to change audio and subtitle names, languages, and flags for all of its files at once, such as setting audio marked und to Japanese. Use the arrows on a track to move it up or down, for example to put Japanese audio before English in every episode. Each row applies to the same ordered track in every episode, using that episode's own track ID and preserving its compression. Only audio tracks you change are rewritten. A new order remuxes the files, while metadata alone is edited in place. Queue changes validates the whole group against the audit before queueing one metadata repair per file. Save changes to defaults to Existing MKV files, updating each file after verification. Choose New .repaired.mkv copies if you prefer separate output files. Follow progress in the job list and audit again after completion to inspect the results.

Fonts in the audit rebuilds the fonts of every audited file without changing a track. Each file gets the font group faces its subtitles need, with real bold and italic faces where lines use them, and a file whose fonts already fit is left alone. With Cut every font to the characters the subtitles use, the font group faces and the fonts that came with the release are cut down to the characters their lines draw, which takes the fonts of a typical fansub episode from tens of MB to a few. Rabbit keeps one full copy of every font it cuts, stored once by its content hash however many episodes carry it, and gives a cut font back the characters later adding, translating, editing or repairing subtitles in Rabbit needs. A font whose licence does not allow cutting, or that no line draws in, stays whole. Rebuilding with the option off makes every cut font whole again. Analyze savings works out the size of the fonts now, after a rebuild that keeps them whole and after one that cuts them, along with the full copies cutting would keep, all in a scratch folder without changing a file. The section shows how much space the full copies take, and Clear deletes them, after which the files play the same but cut fonts can no longer get characters back.

Subtitle timing in the audit checks whether the text subtitles of every audited file follow the speech in its audio. Check timing reads the audio for speech and lines each dialogue track up with it, in a scratch folder without changing a file. A track is only offered a fix when the moved track clearly fits the speech better, so a weak match never moves anything, and shifts under 0.3 s are left alone. A fix can be a shift, a frame rate change such as 25 to 23.976 fps, or a different shift after a cut. When the file has a dialogue track already in sync, the fix is fine-tuned on its lines, because subtitles show slightly before the speech that detection finds. Signs, songs and forced tracks show with on-screen text rather than speech, so they are never measured. They move with the dialogue track of the same language, preferring the one whose title names the same group, and stay unchecked when there is none or the candidates disagree. Commentary and picture subtitles are not checked. Every track that is off is ticked, and an Extra shift in seconds can be added to any listed track. The same change found in several files is named once, which points at the release instead of a single episode. Queueing moves only the chosen tracks, ends lines that now run past the video with it and keeps everything else as it was.

Use Subs from files to add a folder of ASS, SSA or SRT files to the episodes they belong to, such as a fansub translation shipped without video. Choose the folder of videos and the folder of subtitle files, then Pair folders. Files are matched by season and episode number, then by filename order, and every row can be changed or skipped. The language and the credit for the track name are read from names such as .ru.ass or Group [Rus] and can be changed. Each video gets the file as a new full subtitle track, flagged and ordered like a downloaded subtitle, and is replaced in place after the mux.

The fonts the subtitles use are checked against the fonts attached to each video and the fonts inside the subtitle folder or a Fonts folder next to it. Fonts that neither has are listed as missing, since players would draw that text in another font. Upload a font on its row to supply it. A font with another family name is renamed to the name the subtitles use, so a substitute works. Upload several fonts keeps their own names and matches them to the fonts the subtitles use. Each episode gets only the fonts its own file uses, and a font the video already carries under the same name is not attached again. When Replace dialogue font is on for full subtitles, the dialogue fonts are replaced by the font group and are not needed.

Use Font check to find the fonts the subtitles of a folder, or of single videos chosen with Choose videos..., use but the videos do not carry. Every ASS track of every MKV is read in full, so overrides such as \fn inside lines are found too, and each video is read from disk once. Each missing font lists how many videos lack it. Upload... attaches a font to the videos missing it, renamed to the missing name when its family differs. A font in the font library that already has the missing name can be attached as it is. Use another font points the subtitles at it instead, which rewrites the font names in the ASS tracks and remuxes the video. The fonts are listed in groups. Fonts already in the videos add no size. Fonts from the font library are attached to the videos that lack them, with their size shown. A video without a carried font that is pointed at gets the library copy of that family when there is one. For each missing font the carried font most like it is marked closest, judged by its name: another weight of the same family first, then one of the same kind (sans, serif, mono or script), then the font most videos carry. Every font left missing applies one font to the rest in one step, or the closest one to each. Add to library... stores font files in the library, each in a group named after its family, so every later Font check can use them. A video that only gets uploaded fonts is changed in place with mkvpropedit, without a remux. The fixes run as repair jobs, which verify every file before replacing it, and can write *.repaired.mkv copies instead.

Audio Tools

The Audio menu holds two tools that change the audio of existing MKVs without re-encoding their video. Both use the same audio handling as encodes, so a track is kept, ordered, named, flagged and encoded the way an encode with the same settings would do it. Main tracks get no name, other kinds get a label such as Commentary [CR], and the first main track of each language is the default. Opus already at or under its target bitrate is copied instead of encoded again, and nothing is encoded when the audio mode is Copy. Both tools run as jobs that verify the video, subtitles and length of every file before replacing it, and can write *.audio.mkv copies instead.

Use Audio from releases to add the audio languages your videos lack from other releases of them, such as an English dub from a BluRay release added to a WEB encode. Choose a folder of videos, or single videos with Choose videos..., and one or more release folders, or single release files with Add file..., then Read releases. Episodes are matched by season and episode number, then by filename order. A season can be named apart from the episode, as in Show S2 - 02, and a release numbered across seasons, such as Show - 13 for S02E01, is paired by the offset its episodes agree on. When a whole release folder is paired off by some episodes, type its Offset, such as 12 for that release or -1 when every episode got the next one. The offset replaces the automatic pairing of that folder and applies as soon as you press Enter or leave the field. Each episode shows its release as a list of the files in that folder, named by what tells them apart, such as 01. Picking another one pairs that episode by hand, and Shift below then moves every later episode by the same number of files, which fixes a release with an extra recap in the middle. Automatic undoes a pick. A release paired with two episodes is marked on both. A movie is easiest to pair by choosing its release file, since a folder of several releases gives no safe match. The grid has a column per language and kind (main, commentary, audio description, karaoke). Languages and kinds your audio settings keep are picked on their own, the best quality first: lossless, then more channels, then bitrate, or the smallest when that is the codec priority. Anything else can be picked by hand, and a track the video already has is only replaced when you choose Replace with. Check timing compares the audio of one episode with its releases before you queue it, and Check timing of all does it for every episode, two at a time, and counts the ones that did not match.

Each release is lined up with the video by comparing the audio in the video's own language, so a release must carry that language too. A release with only a dub rarely matches and is left out with a warning, while the video keeps its own tracks. A release that only starts at another time has its tracks copied as they are, moved by mkvmerge. One with another frame rate, cut scenes, or audio running more than a second past the video is rebuilt on the video's timeline with FFmpeg. The parts the video lacks are taken out, footage the release lacks is filled with silence, and a PAL speedup is slowed back down by resampling, which also lowers its pitch again. Rebuilt tracks are encoded to Opus. With Apply Rabbit's audio settings on, the default, added tracks are always encoded to Opus and all audio is ordered, named and flagged as an encode would do it. Turn it off to keep lossless copies where the timing allows, with the release's own track names.

Use Audio tracks to give a folder of existing videos the audio your current settings would give an encode, such as older encodes made before you changed the audio settings, or remuxes with large lossless tracks. Choose a folder, or single videos with Choose videos..., then Read tracks. Every kind of track, such as the English 5.1 of every episode, is listed with what your settings do with it and why, for example Not one of your audio languages or At 256 kbps. Its choice of Follow settings, Copy as is, Encode to Opus or Remove applies to that track in every file. Each file shows its tracks in their new order and what changes. Files whose audio is already as the settings want it are left alone.

HDR Metadata

Tools -> HDR Metadata changes the HDR metadata of AV1 MKVs that already exist, for example encodes made before Rabbit wrote HDR10+, without re-encoding them. Choose One file to inspect a video first. Rabbit reads it with retone and shows its HDR formats, colour description, mastering display, MaxCLL and MaxFALL, how many frames carry HDR10+ metadata with its scenes and peak, whether it has Dolby Vision, and the problems retone check finds. Other videos, such as HEVC originals, are described but cannot be changed.

Setting Choices
HDR10+ Keep, measure it from the video, measure it with the scenes of a Dolby Vision original, import it from a JSON file or another AV1 video, or remove it
Dolby Vision Remove it, which leaves HDR10. Shown when the file has Dolby Vision
HDR10 static metadata Keep, set the mastering display (BT.2020 or Display P3 with its luminance) with MaxCLL and MaxFALL taken from HDR10+ or entered, or remove it
Colour tags Tag the video as BT.2020 PQ. Offered only for video without a transfer tag, since tagging SDR as PQ makes it look washed out

Every change becomes a job in the queue with Analyze, HDR10+ when measuring, Write and Replace steps. Rabbit extracts the AV1 track, writes the changes with retone, rebuilds the MKV with MKVToolNix from the new video and every other track, attachment, chapter and tag of the original, and gives the video track its name, language, flags, frame timestamps, default duration and colour elements back. The rebuilt file replaces the original only after its tracks, attachments, chapters, duration and timestamps are verified, so a failed job leaves the file as it was. Measuring decodes the whole video, the other changes take seconds.

An HDR10+ JSON file is uploaded from the browser and can be up to 512 MB, enough for a feature film. It must have one entry per frame of the video, as must the Dolby Vision original. A file with Dolby Vision can only be changed when Dolby Vision is removed as well, because the MKV keeps Dolby Vision's configuration on the video track, which a rebuilt track would lose.

Choose Every MKV in a folder to queue the same changes for a whole folder, optionally with its subfolders, up to 500 files. A folder can keep, measure or remove HDR10+, but not import it, since every file would need its own original. Files the changes do not fit are skipped and keep their content: SDR files when adding HDR10+, files that already have HDR10+ when Leave files that already have HDR10+ as they are is on, and files with nothing to remove. The job shows why it skipped a file.

Film Grain

Tools -> Film Grain changes the film grain synthesis of AV1 MKVs that already exist, without re-encoding them. The player draws AV1 film grain from a few values in every frame header, so changing it leaves the encoded picture and the file size as they are. Use it to take grain out of an encode you do not like, to add a little grain that hides banding in clean or animated video, or to make grain weaker or stronger after watching the result.

Choose One file to inspect a video first. Rabbit reads its film grain with regrain and shows how many table entries it has, the average and strongest luma grain and the chroma grain on the 0 to 255 scale, and the problems regrain check finds. The modal shows how far the reading is: regrain first reads the table out of the video, then describes and checks every entry on all CPU threads. A film with an entry for every frame, as SVT-AV1 writes it, takes under half a minute. Download table saves the video's film grain table, which can be edited and uploaded again or applied to another video.

Film grain What it does
Grain of a chosen strength An ISO or a luma deviation, chroma, and Fine, Medium or Coarse grain, written with regrain generate like the encode's synthetic grain. The ISO is for the video's frame size
Film grain table An uploaded .tbl from regrain, aomenc or Download table, matched to the video by time
Grain measured from the original regrain diff compares the original with the video and puts back the grain the encode lost, optionally just enough to hide banding
Remove it The video plays without grain

An uploaded table is checked with regrain check, and what --fix can correct is corrected, such as comments above the filmgrn1 line or scaling points out of order. A table with errors regrain cannot correct is refused. Static grain pattern gives every frame of a table entry the same pattern, so it stands still instead of moving. Adding grain to a video that has some replaces it. Every change becomes a job with Analyze, Write and Replace steps: Rabbit extracts the AV1 track, writes the grain with regrain, decodes the result with dav1d, and rebuilds and verifies the MKV the same way as the HDR Metadata tool before it replaces the original. A file with Dolby Vision cannot be changed, because a rebuilt video track loses Dolby Vision's configuration. Remove it with HDR Metadata first.

Measuring takes the original the video was encoded from, with the same frames. Rabbit finds the crop and the downscale between them on its own, the way its encodes cut black bars and scale to 1080p, and takes out the video's own grain first so it does not count as grain the encode lost. Quick (30 s) measures a representative half minute and gives the whole video that grain, Full video gives every scene its own, and Scan reads only part of each 48 frames to go faster. Luma grain measured this way matched the grain measured against the clean source within 1% on 360p and 720p test videos from CRF 20 to 45. Chroma is measured only with Measure chroma grain too, because colour detail the encode lost counts as chroma grain: a source without any chroma grain measured 0.9 to 5.4 on the 0 to 255 scale from CRF 25 to 45. A folder cannot measure, since every file needs its own original.

Choose Every MKV in a folder to queue the same change for a whole folder, optionally with its subfolders, up to 500 files. Files without grain are skipped when removing it, and files that already have grain when Leave files that already have film grain as they are is on.

Denoising

Denoising runs in the Prepare step, after crop / downscale / deband. The Denoise setting picks the level (off, auto, light, medium, heavy) and Video handling -> Denoiser picks what does the work:

Engine Runs on Notes
rabbit-denoise GPU, or CPU rabbit-denoise film grain denoiser. Measures the grain of each scene at every frequency and brightness and removes it while keeping detail. Made to pair with Restore denoised grain.
avd-nl4d GPU only av-denoise NL4D spatio-temporal denoiser. The default. Best noise removal and detail retention, and the slowest.
avd-nlmeans-hq GPU only av-denoise NLMeans-HQ. Measures the noise level itself and keeps more detail than plain NLMeans.
avd-nlmeans GPU only av-denoise fast NLMeans. Hand-set strength, light denoising only.
nlmeans CPU, OpenCL or Vulkan FFmpeg's nlmeans. Falls back to CPU when no GPU backend works.

The av-denoise engines use Vulkan (or CUDA when an NVIDIA GPU is passed through to the container) and have no CPU fallback: if av-denoise cannot use the selected device, the job fails instead of encoding without denoising. Vulkan works with the /dev/dri device mapping from the stock docker-compose.yml on AMD and Intel GPUs.

rabbit-denoise uses Vulkan through the same /dev/dri mapping on AMD and Intel GPUs, and picks the GPU itself. Its Device setting is auto by default: without a usable GPU the job runs on the CPU, about 20 times slower, and the log says so. gpu fails the job before any work starts instead, and cpu always uses the CPU. Rabbit checks the GPU by denoising one small frame. The output keeps the source's bit depth and chroma layout. The container downloads the checksum-verified rabbit-denoise v0.1.11 release.

Advanced Settings -> Rabbit Denoise Settings holds all of its options, grouped as in its own help: the device, the strength (scale, chroma scale, threshold and first stage only), matching (step, temporal radius, search radius, matches per frame, minimum shift, group size and both match limits), motion (motion fields, motion range and both follow radii) and grain measurement (single frame cap and luma guided chroma). Every value has its default and range shown and is limited to what rabbit-denoise accepts. The defaults are the best quality measured, and only changed values are passed, so a newer rabbit-denoise brings its own improved defaults. Reset Rabbit Denoise settings to defaults undoes every change. The benchmark always times it on the GPU.

The engines differ in how the level is used:

  • rabbit-denoise, avd-nl4d (the default denoiser) and avd-nlmeans-hq measure the noise in each scene themselves and set their own strength. light, medium and heavy therefore denoise identically, and the level only decides whether a scene is denoised. Each has a single expert scale in Advanced Settings (--sigma-scale from 0.25 to 4, --lambda-ht-scale and --hq-sigma-scale, default 1.0) that applies at every level.
  • avd-nlmeans needs its strength set by hand, like FFmpeg nlmeans, so light / medium / heavy map to --strength 1.0 / 1.5 / 2.0, the same values as the FFmpeg nlmeans levels. The value applies to both luma and chroma.

The two NLMeans engines also have an AVD Speed Preset (veryfast to veryslow) that trades speed for quality. NL4D always runs at its default preset.

Auto denoise works with every engine. Scenes are classified by noise, by bitrate or by grain against your thresholds, then each classified range is denoised by the selected engine and untouched ranges are passed through. With rabbit-denoise, avd-nl4d and avd-nlmeans-hq every scene at or above the light threshold is denoised the same way, and neighbouring ranges are denoised as one stretch. rabbit-denoise takes the whole plan in one run with --denoise-frames: it denoises the planned frames, passes the others through unchanged, and the video is neither cut into segments nor joined again. A rabbit-denoise older than 0.1.9 gets each range on its own.

Advanced Settings -> Auto Denoise Metric -> grain classifies scenes by the grain rabbit-denoise analyze measures in each of its own scenes: the deviation of the luma grain on a 0 to 255 scale, whatever the bit depth. It measures the grain in single frames and between frames and takes the smaller, so texture and motion do not read as grain, and clean scenes are copied untouched. Every scene keeps its own reading, however short, so the scenes denoised are exactly those rabbit-denoise --min-grain would choose. Clean anime reads below 1 and grainy openings and endings 3 and more, so the default light threshold of 2 denoises only the grainy scenes, such as a grainy opening in an otherwise clean episode. Anime with film grain throughout, such as older cel animation scanned from film, reads 1.5 to 2.5 from scene to scene, and is better denoised whole or with light at 1. Medium (3) and heavy (5) only matter for plain NLMeans. The grain metric works with every denoiser, as rabbit-denoise only measures. The chart of a queued job shows every scene's grain against the thresholds. The analysis reads the whole video at 80 to 90 fps at 1080p, about 7 minutes for a 24 minute episode.

With Restore denoised grain, regrain gives every scene of the video its own grain. A sampled Scan places a scene change only within some frames of where it is, so at the start and end of a denoised range a few frames can get the grain of the range next to them. Scan -> Every frame puts every change on its frame.

The first run with a new combination of av-denoise settings compiles GPU kernels, which takes a few seconds. They are cached in /config/vapoursynth/av-denoise-cache (override with AV_DENOISE_COMPILATION_CACHE).

The Benchmark dialog measures every engine at all three levels. Tick the engines you want to compare before running it.

VapourSynth Filters

Rabbit Encoder ships with a VapourSynth filter system that lets you stack arbitrary preprocessing passes in front of the main encode. Each filter runs as its own vspipe pass.

Stock presets

Built-in presets live under /app/vapoursynth/presets inside the container and are namespaced as stock:<id>:

Preset ID Name Description
stock:finedehalo FineDehalo Reduces dark/bright halos around high-contrast edges. Best for anime and DVD upscales.
stock:dehalo_alpha Dehalo Alpha Classic halo reduction via vs-jetpack's dehalo_alpha.
stock:f3k_deband F3K Deband Removes color banding in smooth gradients while preserving detail. Wraps vszip.Deband (modern f3kdb successor) and can re-grain after debanding.
stock:source_bitdepth Source bit depth Choose 8, 10, 12, or 16 bits for the filter pass. By default, sources below 10 bits are raised to 10 bits. Turn on Convert sources above target to reduce sources with more bits. Reductions use ordered dithering. The preset does not trim frames or change frame rate.

Every preset declares its own levels (typically light, medium, heavy) and a set of tunable parameters. You can pick a level per job and override individual params per-level in the dashboard's Advanced Settings -> VapourSynth Filters panel.

The source bit depth preset changes the video sent to later filters. The final AV1 encode converts video to 10 bits. Choosing 8 or 12 bits for this preset does not make an output file with that bit depth. Reducing bit depth discards precision even if a later filter raises it again.

User custom presets

To add your own filter, drop a matching pair of files into the host-mounted user directory (./vapoursynth-user/presets by default, exposed inside the container at /config/vapoursynth/presets):

myfilter.vpy    # the VapourSynth script
myfilter.json   # the manifest (id, levels, params, defaults)

Both files must share the same stem. The .vpy script reads its input path from the SRC argument and any tunable parameter via rabbit_vs.arg_int / arg_float / arg_str / arg_bool. See Examples.

User presets are namespaced as user:<id> and override nothing (stock and user presets coexist). After editing or adding presets, click Reload presets in the Advanced Settings panel (or POST /api/vs-presets/reload). No container restart is required.

Per-job behavior

The VapourSynth chain is stored per job, so different files in the same queue can use different filter stacks. Each entry has a level (or "off" to disable without removing it from the chain) and a per-level parameter map, letting you switch intensity without losing your tweaks at other levels. The active chain and its resolved parameters are stored in the output MKV's versioned SETTINGS code (for example, RE1|...) so the configuration can be reproduced later.

Subtitle Fonts

Subtitles are styled per font group. A group is a folder under the user fonts directory (/config/fonts) containing one or more font files plus an optional metadata.json. A single group can supply different faces for different writing systems. For example a Latin-only face like Trebuchet MS together with a CJK face like Noto Sans JP. This lets one group cover scripts that no single font does.

Noto Sans and Noto Serif ship with Rabbit Encoder and are seeded into /config/fonts on first start if they are not already present, so they are editable in place.

metadata.json

{
	"label": "Anime old",
	"faces": {
		"trebuchet.ttf": { "keys": ["latin"] },
		"noto_jp.ttf": { "keys": ["japanese", "jpn"] },
	},
	"style": {
		// group-global appearance (1080p px)
		"fontSize": 80,
		"outline": 4,
		"shadow": 1.5,
		"marginV": 50,
		"marginL": 135,
		"marginR": 135,
		"alignment": 2,
		"bold": false,
		"primaryColour": "&H00FFFFFF",
		"outlineColour": "&H00000000",
		"backColour": "&H80000000",
	},
	"overrides": {
		"jpn": { "fontSize": 72, "marginV": 60, "fontAxes": { "wght": 600 } },
		"japanese": { "fontSize": 74 },
		"latin": { "fontAxes": {} },
	},
}

For each subtitle track, the face and the appearance are both resolved by the track's language tag and the detected writing system, in this order:

  1. a keys/overrides entry matching the language code (e.g. jpn, de),
  2. then the writing system (e.g. latin, japanese, cjk),
  3. then the group-global style,
  4. then the built-in defaults.

The family name written into the ASS is taken from the resolved face, so style/overrides only carry appearance (size, outline, shadow, margins, colours, alignment, bold, and variable-font axes). Because axes are face-specific, they belong in the per-key override for that face (a static Latin face has none, a variable CJK face can pin wght).

You can manage fonts entirely from the dashboard (Tools -> Fonts, or "Manage fonts..." next to the subtitle style). All fonts lists every face of every group with a preview in its own font, its file, size and group. Search it by family, file or group, and remove faces from it. Removing the last face of a group removes the group too. Upload fonts... there stores each font in a group named after its family, where Font check and the subtitle style can use it. Below, you can create, rename, and delete groups, edit each face's keys, and remove faces. All changes are written to /config/fonts and the registry is reloaded automatically. Renaming a group also repoints the fontGroup reference in your default settings and in every queued job, so existing selections aren't orphaned.

If you mount a host font directory read-only, those fonts become selectable in the import dropdown: pick a font, choose a writing system (Latin, Japanese, Cyrillic, ...) or type extra language keys, and click Import to copy it into the selected group with those keys. The host mount is never written to. Import always copies into /config/fonts.

Upload fonts from this computer stores font files from the browser in the selected group the same way, with a writing system and extra keys.

You can still edit a group's style from the dashboard ("Subtitle font & style..."), choosing a scope of "Group global" or a specific script/language. Saving writes back to the group's metadata.json. Click Reload fonts (or POST /api/fonts/reload) after hand-editing files on disk.

Variable-font axes are instanced into static fonts before attachment. When a retained source font already uses the selected internal family name, Rabbit Encoder picks the first free numbered alias (Noto Sans → Noto Sans 2 → Noto Sans 3), written consistently into the static font's internal metadata and the ASS styles it rewrites. The attachment filename is slugified (Noto Sans 2 -> noto_sans_2.ttf). The internal family name keeps its spaces, so rendering is unaffected.

Lines drawn in bold or italic, through their style or \b1 and \i1, get real faces of the same family next to the face the style names: noto_sans_bold.ttf, noto_sans_italic.ttf and noto_sans_bold_italic.ttf. Bold is the group's weight plus 300, up to the font's heaviest. Italic comes from an italic font of the same family in the group, such as the latin-italic.ttf that Noto Sans and Noto Serif ship with. Scripts with no italic font, such as Arabic or Japanese, are slanted by the player. A shipped face added in a later version is copied into an existing group once, and stays deleted if you remove it.

With Attach only the characters the subtitles use enabled, each of these faces keeps only the characters its lines draw, plus ASCII, Latin-1 and common punctuation, which brings a face from about 650 KB to about 40 KB. A face cut down this way cannot show new characters typed into the subtitles in another tool. A file cut once stays cut through later passes. To cut existing files, or make them whole again, use Fonts in the folder audit. Every face records how it was made, so adding, translating, editing, repairing or removing subtitles in Rabbit Encoder makes the faces again for the text the file then has.

With Remove unused fonts enabled, source attachments are kept only when surviving untouched ASS styles or inline \fn overrides still reference them.

Output Naming

Files are named following the pattern:

{Title} [Source-Resolution][Opus Layout][AV1]-{ORGANIZATION}.mkv

For example:

Blue Exorcist (2011) - S01E01 - The Devil Resides in Human Souls [Bluray-1080p][Opus 2.0][AV1]-RabbitCompany.mkv

Source tags are detected from the input filename: Bluray, WEBDL, WEBRip, HDTV, DVD, SDTV, CAM. Files with REMUX in the name are tagged as Bluray.

Disc Naming

Episodes, movies and extras from Blu-ray discs are named by formats you set in docker-compose.yml, with placeholders written like Sonarr's and Radarr's. The encode adds its tags after the name as above.

Variable Default
SERIES_FOLDER_FORMAT {Series TitleYear}
SEASON_FOLDER_FORMAT Season {season:00}
SPECIALS_FOLDER_FORMAT Specials
EPISODE_FORMAT {Series TitleYear} - S{season:00}E{episode:00} - {Episode Title}
MOVIE_FOLDER_FORMAT {Movie Title} ({Release Year})
MOVIE_FORMAT {Movie Title} ({Release Year}) - {Edition Tags}

Series placeholders are {Series Title}, {Series TitleYear}, {Series Year}, {Year}, {ImdbId}, {TvdbId}, {TmdbId}, {season}, {episode} and {Episode Title}. Numbers are padded with {season:00} or {episode:000}. Movie placeholders are {Movie Title}, {Movie TitleYear}, {Release Year}, {Year}, {ImdbId}, {TmdbId} and {Edition Tags}. Placeholders are not case sensitive, and ones that do not exist are logged at startup and left empty.

A placeholder without a value takes the brackets around it along, and the dash before it, so [imdbid-{ImdbId}] and - {Edition Tags} disappear when there is no ID or edition. Characters that cannot be in file names are replaced as Sonarr replaces them. Extras always go to an extras folder inside the series or movie folder, where Jellyfin and Plex look for them.

For the names of the TRaSH guides:

- MOVIE_FOLDER_FORMAT={Movie Title} ({Release Year}) [imdbid-{ImdbId}]
- MOVIE_FORMAT={Movie Title} ({Release Year}) [imdbid-{ImdbId}] - {Edition Tags}
- SERIES_FOLDER_FORMAT={Series TitleYear} [tvdbid-{TvdbId}]

which gives

Your Name. (2016) [imdbid-tt5311514]/Your Name. (2016) [imdbid-tt5311514] [Bluray-1080p][Opus 5.1][AV1]-RabbitCompany.mkv

Settings Codes

Each encoded file includes a SETTINGS tag. This is a compact, versioned string that records the exact resolved settings used for the encode, such as RE1|c~q=h,ds=1|dn~m=m,s=4.5,p=5,r=13.

The settings code stores the actual underlying values, including every nlmeans and gradfun parameter, each active VapourSynth filter in chain order, and all related filter parameters. This makes it possible to reproduce a setup even if presets have changed since the original encode.

To restore settings from a code, paste it into the Settings Code field in either the Default Settings panel or the per-job settings panel. Codes are portable across machines because GPU-specific choices, such as the denoise backend and GPU device, are intentionally left out. This prevents an imported code from overwriting the target machine’s hardware configuration.

When an AV1 film grain table was used for the encode, the code carries fg=1 in its core section, for example RE1|c~fg=1. The flag is left out when no table was used. It only records that a table was used. The table itself and its path are not stored, so importing a code never changes a job's film grain table.

The format is versioned with an RE<n> prefix, so older codes continue to work as the format evolves. Fields that match the shipped baseline are omitted, which means a stock-default configuration is represented simply as RE1.

Supported Input Formats

.mp4, .mkv, .avi, .webm, .flv, .ts, .mov

API Endpoints

Method Endpoint Description
GET /api/jobs List all jobs
GET /api/jobs/:id Get job details
PATCH /api/jobs/:id Update job settings (queued only)
DELETE /api/jobs/:id Remove a job
POST /api/jobs/:id/retry Retry a failed job
POST /api/jobs/:id/cancel Cancel an actively encoding job
POST /api/jobs/:id/move Move a queued job in the queue (direction: up, down, top, bottom)
POST /api/jobs/:id/import-code Apply a settings code to a queued job, replacing its settings
POST /api/jobs/reorder Set the entire queue order from a JSON { ids: [...] } body
GET /api/jobs/:id/audio-preview Preview audio reorder/filter/dedup for a job
GET /api/jobs/:id/subtitle-preview Preview subtitle reorder/rename for a job
GET /api/jobs/:id/subtitle-tracks List source subtitle tracks (for the translate-only source picker)
GET /api/jobs/:id/mediainfo Run mediainfo on the source file and return the report
GET /api/jobs/:id/bitrate-analysis Bitrate + noise/bitrate/grain-metric scene data, cached (?refresh=1 to re-run)
GET /api/jobs/:id/measure The finished encode's quality measurement against its source, if any
POST /api/jobs/:id/measure Score the encode against its source. Body: interval, metrics, withGrain
DELETE /api/jobs/:id/measure Cancel the running measurement
GET /api/bitrate-analysis?path= Bitrate over time of any video in the media folders (?refresh=1 to re-run)
GET /api/jobs/:id/live Steps and Rabbit Boost telemetry, telemetry only when newer than ?version=
GET /api/jobs/:id/preview Get preview-encode state for a job (idle, running, or completed samples)
POST /api/jobs/:id/preview Start a preview encode. The optional body configures sample count and duration
DELETE /api/jobs/:id/preview Cancel a running preview, or clear completed preview artifacts
GET /api/jobs/:id/preview/sample/:index/:kind Fetch source/encode clips, comparison PNGs, VS stills, or prepare-stage stills
GET /api/repair/inspect Inspect target/source MKV subtitle tracks for the repair editor
GET /api/repair/roots List configured media roots for the repair file picker
GET /api/repair/browse Browse MKV files within an allowed repair-picker root
POST /api/repair/audit Group a folder or completed job outputs by audio/subtitle metadata
POST /api/repair/audit/group Validate an audited group and queue subtitle metadata edits for all its files
POST /api/repair/audit/subtitle-timing Check how well the text subtitles of up to 4 audited files follow their audio
POST /api/repair/audit/sync-subtitles Queue repairs that move chosen subtitle tracks to fit the audio
POST /api/repair/jobs Validate and queue a subtitle-only stream-copy repair plan
GET /api/config Get default settings
PATCH /api/config Update default settings
POST /api/config/reset Reset default settings
POST /api/config/import-code Import a settings code as the new default settings
POST /api/settings/encode Encode a settings object into a shareable settings code ({ code })
POST /api/settings/decode Decode a settings code back into a settings object ({ settings })
GET /api/library List configured library root directories
GET /api/library/browse Browse a library folder (?path=/data/library/Animes)
POST /api/library/encode Queue all videos in a folder for in-place encoding
GET /api/queue Get queue state (paused or running)
POST /api/queue/pause Pause encoding. Stops current encode and preserves the queue
POST /api/queue/resume Resume encoding from where it was paused
GET /api/fonts List font groups and their faces
POST /api/fonts/reload Rescan the user fonts directory and reload the registry
GET /api/fonts/:label/style Get a group's style (group-global + per-language/script overrides)
PUT /api/fonts/:label/style Save a group's style into its metadata.json (user fonts dir)
PUT /api/fonts/:label/style Save a group's style into its metadata.json (user fonts dir)
GET /api/system-fonts List importable fonts under the read-only host font directories
POST /api/fonts/groups Create a new font group folder under the user fonts dir ({ label })
PATCH /api/fonts/groups/:label Rename a group (folder + label) and repoint defaults and all jobs
DELETE /api/fonts/groups/:label Delete a font group folder and its contents
POST /api/fonts/groups/:label/faces Import a host font into the group with writing-system / language keys
POST /api/fonts/upload Store an uploaded font in ?group=, or a group named after its family
PATCH /api/fonts/groups/:label/faces/:file Update a face's keys / family override in the group's metadata.json
DELETE /api/fonts/groups/:label/faces/:file Remove a font file from the group and its metadata entry
POST /api/subsearch/search Start a subtitle search for one video: download, sync and rate per language
GET /api/subsearch/search/:id Progress and result of a subtitle search
DELETE /api/subsearch/search/:id Cancel a subtitle search or drop its result
POST /api/subsearch/apply Queue the chosen subtitles as new tracks, or save them next to the video
POST /api/subsearch/test Check an OpenSubtitles login and report the download quota
POST /api/subsearch/test-subdl Check a SubDL API key
POST /api/subtitle-files/scan Pair a folder of videos with a folder of subtitle files and report their fonts
POST /api/subtitle-files/font Upload a font, renamed to ?name= when its family differs
POST /api/subtitle-files/jobs Queue one subtitle add per paired video, with the fonts each file uses
POST /api/font-check/scan Start reading a folder's subtitles for fonts its videos do not carry
GET /api/font-check/scan/:id Progress and result of a font check
DELETE /api/font-check/scan/:id Cancel a font check or drop its result
POST /api/font-check/jobs Queue one repair per video, attaching uploaded or library fonts and mapping
POST /api/release-subs/scan Pair a folder of encodes with folders of releases and plan the subtitles
POST /api/release-subs/check Compare the audio of one episode with its releases
POST /api/release-subs/jobs Queue one repair per episode, retiming the chosen subtitles by audio
POST /api/release-audio/scan Pair a folder of videos with folders of releases and plan the audio
POST /api/release-audio/jobs Queue one job per episode, syncing the chosen audio tracks by audio
POST /api/audio-tracks/scan Plan the audio of a folder as the settings and per-track choices say
POST /api/audio-tracks/jobs Queue one job per file whose audio changes
POST /api/ocr/test Read a sample subtitle image with the OCR model
GET /api/ocr/inspect Subtitle tracks of an MKV, marking picture tracks without a text version
POST /api/ocr/jobs Queue OCR of chosen picture subtitle tracks of one MKV
POST /api/ocr/pairs Match a folder of encodes with a folder of their originals
POST /api/ocr/batch Queue OCR for every MKV in a folder
GET /api/system Current system resource usage (CPU, RAM, temp-partition disk, network, GPU)
GET /api/opencl-devices List available OpenCL devices
GET /api/vulkan-devices List available Vulkan devices
GET /api/avd-devices List GPU devices usable by av-denoise
GET /api/benchmark Get current benchmark state
POST /api/benchmark Start a denoise benchmark run
DELETE /api/benchmark Cancel a running benchmark
GET /api/vs-presets List all VapourSynth presets (stock + user) with their manifests
POST /api/vs-presets/reload Rescan stock and user preset directories from disk and reload the registry
GET /api/vs-presets/:id/default-entry Build a fresh filter-chain entry for :id, pre-filled with manifest defaults

License

GPL-3.0

About

Automated AV1 video transcoding pipeline powered by Auto-Boost-Essential and Opus audio encoding, with a real-time web dashboard.

Resources

Stars

23 stars

Watchers

2 watching

Forks

Releases

Contributors

Languages