Above: two commercial blocks taken out of a 3:45 recording — 133.91 s copied bit-for-bit, 0.57 s re-encoded. 17 frames out of 6743 were touched at all.
The clip is a synthetic test recording, built by
tests/make_demo_media.sh
from nothing but ffmpeg's own sources: colour cards, a fake station logo, and a
15-second commercial grid.
Cutting video normally means decoding and re-encoding the whole file. SmartCut re-encodes only the partial GOPs that the cut points fall inside, and copies everything else bit-for-bit.
... I ....... I=========================I ....... I ...
^t_in ^k_first ^k_term ^t_out
|<-head->|<--------- body --------->|<-tail->|
re-encode stream copy re-encode
On real terrestrial recordings over 99% of the output is a lossless copy. A 22-minute, 5-range export driven by the automatic commercial detector came out bit-identical across all 40589 frames.
- Smart rendering — H.264 / HEVC / MPEG-2 / MPEG-4 Part 2. Cutting exactly
on an access point re-encodes nothing at all. The audio works the same
way (
--audio-mode smart, the default): only the AAC frames a boundary falls inside are re-encoded, so nothing from the far side of a cut is heard -- 4 frames out of 5606, measured, and none at all when the seam falls in silence, as a commercial cut does -- where the output is byte-identical to a copy. And it stays MPEG-2 AAC, which left to FFmpeg it would not: the ADTS headers are written here so a seam is not one MPEG-4 frame in an MPEG-2 stream. - 5.1 folded down when it has to be (
--audio-channels 2, or the output settings screen). A surround recording that has to play somewhere that will not have it is the one case where copying the audio is not the answer, so this re-encodes the track -- with libav's own downmix coefficients, and with the ADTS headers rewritten to say stereo, which is where a transport stream states its channel count. - Captions and programme information come through (writing a .ts). The ARIB STD-B24 caption stream the broadcast sends is carried across, shifted by exactly what the pictures were shifted by -- a caption statement is whole inside one packet, so there is nothing at a seam to re-encode and it arrives byte for byte. So do the programme on now and the one after (EIT), the station name (SDT) and the broadcast clock (TOT): the muxer writes a description of the streams and stops there, so the finished file is walked once more and the recording's own tables are put back. Every stream returns to the PID it arrived on, which is where a tool built around broadcast recordings looks for it. Superimposed text and the data broadcast cannot come -- the first arrives with no time on its packets and the second is a carousel of sections rather than a stream -- and the tool says so rather than dropping them quietly.
- Multi-audio broadcasts are read, and both tracks are written. A bilingual broadcast sends its two sound tracks on separate PIDs, and only one of them used to be looked at. Smart rendering runs on each track independently: two tracks have their frames at different instants and drift by different amounts, so one track's answer cannot stand in for the other's. The one you do not want is switched off in the cut editor's track menu.
- Automatic commercial-boundary detection — reads the junction marks the broadcast puts in its own caption stream, runs of silence, and the presence of the station logo, and reports the runs that land on the 15-second grid. Boundaries snap to access points, so cutting commercials stays entirely lossless.
- Cut-editing GUI — film strip, scene detection, scroll search, and preview playback with audio. What you see is always the edited timeline.
- A clip list, and a batch behind it — drop a night's recordings on the
input screen and they are read in the background, each leaving its seek
index on disc;
Ctrl+AthenCtrl+Dsets a commercial detection running on all of them. Reading, detecting and cutting run at the same time: the batch does not stop for the editor, and a clip the batch has not reached can be opened anyway. The whole list — recordings, cuts and output settings — saves as a project withCtrl+Sand comes back the next evening. The cut editor opens on one clip in a window of its own and closes with OK; cuts stay with the clip, so you can work through the list and then write the lot out in one go. A clip can be duplicated, cuts and marks and all — one recording cut two ways, both written out, sat side by side in the list. What the output screen shows is not a poster frame but the frames that will actually be re-encoded — everything else is copied byte for byte. - A seek index — the two passes that used to be repeated on every open (walking the packets for the access points, decoding the key pictures for the thumbnail track) are done once and written down. Opening a half-hour recording a second time goes from 18 seconds to 0.1. The index also carries the byte offset of every access point, which takes the guesswork out of seeking.
- Proxy editing (off by default;
SMARTCUT_PROXY=1) — a small stand-in is built from the recording, and the preview, the film strip and playback read from it. It carries the recording's own timestamps and access points, so cutting still works from the recording itself. It is for material where decoding a single picture is itself too slow to scrub, which broadcast 1440x1080 MPEG-2 is not. - Built for broadcast material — interlacing is preserved, 2:3 pulldown is
handled on a field-level timeline, and dropped frames, non-zero
start_time, and ARIB ADTS layout are all accounted for. - Output containers — MPEG-TS / MP4 / Matroska, defaulting to the same container and directory as the input.
- English or Japanese — the interface follows whatever the machine is set to, and Preferences overrides it. The change lands at once, in both windows, and is remembered for next time.
Codec and track-layout limits apply; see known limitations.
Grab a build from Releases. Only the .deb asks for FFmpeg on the system — every other build bundles it.
| Linux | SmartCut_0.3.1_amd64.AppImage, or SmartCut-0.3.1-linux-x86_64.tar.gz (unpack it, run ./smartcut). Both carry FFmpeg and need glibc 2.39+ (Ubuntu 24.04 / Debian 13 / Fedora 40 or newer) |
| Linux (deb) | smartcut_0.3.1_amd64.deb — sudo apt install ./smartcut_0.3.1_amd64.deb. 2.6MB, because it links the system FFmpeg 7.1 instead of carrying one; that means Debian 13 / Ubuntu 25.04 or newer. Installs the GUI as smartcut and the cutter as smartcut-cli |
| Windows | SmartCut_0.3.1_x64-setup.exe (installer) or smartcut-portable-x64-0.3.1.zip (unzip and run). x64 only; needs the WebView2 runtime |
To build it yourself, see Building and development.
Drop a file onto the GUI, or use the command line:
smartcut input.ts --keep 5.3-12.7 -o out.ts # keep the given range
smartcut input.ts --cut 8.0-20.0 -o out.ts # drop the given range
smartcut input.ts --analyze # show the plan, write nothing
smartcut input.ts --analyze --detect-cm --logo # commercial candidates
smartcut input.ts --analyze --scenes # scene changes| Option | Meaning |
|---|---|
--keep START-END / --cut START-END |
Ranges; repeatable. 1:23:45.6 form also accepted |
--audio-mode smart|copy|reencode |
smart (default) re-encodes only the frames a boundary falls inside, so nothing from the far side of a cut is heard -- and nothing at all when the seam falls in silence. copy is lossless to the byte; reencode is sample-accurate |
--audio-channels N |
Channels to write, 1..8. Anything but the recording's own is a downmix -- 5.1 folded to stereo for the players that make a mess of surround -- and there is no copying through one, so it re-encodes the whole track whatever --audio-mode says |
--audio-bitrate RATE |
Bits per second for re-encoded audio, 192k or 192000. Left out, it follows the recording, and comes down with the channel count when there is a fold |
--aac auto|mpeg2|mpeg4 |
Which AAC the frames this tool writes announce themselves as. auto follows the recording, which for a broadcast means MPEG-2 AAC |
--index scan|container |
How access points are indexed. container is faster but unavailable for TS |
--seek-index PATH |
Where to keep the seek index. Written on the first run, read on the next, which skips the walk over the packets |
--detect-cm / --logo / --scenes |
Commercial candidates, logo assist, scene detection |
--drop-stream INDEX |
Leave one of the recording's streams out of the output; repeatable. The same thing the cut editor's Tracks menu does |
--tables partial|broadcast|muxer |
How a .ts describes itself. The default partial writes a partial transport stream (one SIT, per DVB EN 300 468 Annex C / ARIB TR-B15); broadcast puts the recording's own PMT, SDT, EIT and TOT back; muxer leaves the muxer's own tables standing (once --no-tables) |
--no-open-gop |
Never start a copy at an open GOP |
-o OUTPUT |
Output path; the extension picks the container |
Where a night's worth of recordings goes. Each one is read in the background
as it arrives, leaving a seek index behind; Ctrl+A then Ctrl+D runs
commercial detection over everything selected. The row carries what is known
about that recording — length, resolution, the commercial blocks found, the cuts
made, the length it will be written at. Cuts live with the clip, so the list can
be cut through one at a time and written out in one go. The cut editor opens
on one clip in a window of its own and comes back with OK.
The SmartCut button in the corner of this screen opens Preferences, where the interface language is set: English, Japanese, or whatever the machine is set to, which is the default. The change takes effect at once — both windows — and is remembered for next time.
About SmartCut, in the same menu, prints the versions: the program's and the cutting engine's, the FFmpeg libraries this process actually loaded and what they are licensed under, and the platform. It is the one part of either window whose text can be selected, so it can go straight into a bug report.
The same menu saves and opens projects (Ctrl+S / Ctrl+O). What is written
is the list itself — the paths, the cuts and marks made in each recording, and the
output settings. Nothing that can be worked out again goes in: the seek indexes and
the commercial detections live in the program's own cache directory, so opening a
project simply reads them again. That keeps a .scproj at a few hundred bytes, and it
means a project still opens on another machine, or after the caches have been
cleared. A .scproj dropped on the window opens too, as does
smartcut friday-night.scproj.
The title bar names the project that is open and puts a * in front of it
while there is work that is not on disc — and closing the window on that work
asks first. Being on disc is compared, not tracked: cancelling out of the
editor, or removing a clip that was just added, takes the * off again by
itself.
Cuts are subtractive: the timeline you see is the recording minus the cuts, never the original. A cut region does not turn grey — it disappears. The seek bar shrinks, the film strip closes over the hole, and the frame counter counts the length that will actually be written. All that is left of a cut is a red vertical line at the join.
The status line at the bottom is the plan the engine will execute, before you commit to it: which ranges get copied, which get re-encoded, and how many frames that is.
Every page is available in English and in Japanese; the switch is at the top of each one. The part worth reading first is the pitfalls — the eight reasons why "just cut on GOP boundaries and concatenate" does not work, in the order they were hit.
| Algorithm and pitfalls | How the cut is split, and the eight traps |
| Rust core | Timestamps, mixed SPS/PPS, audio boundaries |
| Validation and limits | Frame-hash verification, real-material results |
| GUI | Editor, the seek index, thumbnail track, scene detection, playback, the proxy |
| Commercial detection | Silence plus logo, the 15-second grid, avoiding false positives |
| Broadcast workflow compatibility | PID layout, ADTS, L-SMASH / DGIndex |
| Building ・ Distribution | How to build and ship it |
| BDMV / BDAV ・ Design notes | Research and design decisions |
rust/ Rust core (smartcut_core) and CLI <- the real implementation
gui/ Tauri v2 + vanilla JS GUI
smartcut/ Python reference implementation <- test oracle
tests/ 11 end-to-end suites, 112 checks
docs/ Documentation
The Python implementation is kept as the reference implementation and test
oracle that pinned down the algorithm and its pitfalls. It shares the same
frame-hash verification with the Rust core, and tests/run_tests.sh and
tests/run_rust_tests.sh report identical lossless ratios.
x264 and x265 are GPL, and linking against them makes the whole application GPL. Re-encoding can also be switched to a hardware encoder (NVENC / QSV / VideoToolbox / AMF). Patent licensing for H.264 / HEVC needs separate consideration for commercial distribution.