bfvt 0.1.0

Black-frame video trimmer: detect long black + silent stretches and remove all but the first moment of each.

bfvt — black-frame video trimmer

Scan a video, find long stretches that are simultaneously black and silent, and remove all but the first moment of each — then write a re-encoded, trimmed file. Built for chopping the dead air out of recordings: the leader/trailer black, the gap between segments, the "I forgot to stop the screen recorder" tail.

Tiny stray bright pixels (codec ringing, a dead-pixel speck, a faint watermark) are still treated as black, so a frame doesn't get disqualified just because a handful of pixels aren't perfectly #000000.

How it decides what to cut

1. Black detection (OpenCV). Frames are sampled (default 10 fps). A frame counts as black when the fraction of pixels brighter than a luma threshold is at or below a small ratio (default 0.2%) — that ratio is the "ignore stray pixels" knob.
2. Silence detection (ffmpeg silencedetect). Audio below a noise floor (default −30 dB) is silence. Videos with no audio track are treated as wholly silent, so black detection alone drives the cut.
3. Interval algebra (pure, deterministic). The black intervals are intersected with the silent intervals. Each overlapping run that lasts at least --min-duration (default 5 s) is trimmed, keeping the first --keep-head seconds (default 0.25 s) so a hard chapter cut doesn't become a jarring jump. Everything else in the run is removed.
4. Re-encode (ffmpeg). The kept segments are concatenated via a trim/atrim + concat filtergraph and re-encoded (libx264 + AAC).

Given the same input and flags, the output is deterministic.

Install

pip install .
# or, for development + tests:
pip install -e ".[test]"

Requires ffmpeg and ffprobe on PATH. OpenCV is pulled in via opencv-python-headless.

Usage

# Trim, writing alongside the input as <name>.trimmed.mp4
bfvt input.mp4

# Choose the output and the threshold
bfvt input.mp4 -o clean.mp4 --min-duration 5 --keep-head 0.25

# See what WOULD be cut without re-encoding
bfvt input.mp4 --dry-run

# Trim on darkness alone (ignore audio)
bfvt input.mp4 --ignore-audio

Key options

Flag	Default	Meaning
--min-duration SEC	5.0	Minimum black+silent run before anything is trimmed
--keep-head SEC	0.25	Seconds at the start of each run to preserve
--luma-threshold N	32	0–255 luma above which a pixel is "bright"
--bright-ratio R	0.002	Max fraction of bright pixels for a frame to still be black
--sample-fps FPS	10	Frame sampling rate for black detection
--noise-db DB	-30	Audio below this dB is silence
--ignore-audio	off	Trim on darkness only
--crf / --preset	18 / medium	libx264 quality/speed
--dry-run	off	Report cuts, write nothing

Library

from bfvt import TrimConfig, analyze, run

cfg = TrimConfig(min_duration=5.0, keep_head=0.25)
res = analyze("input.mp4", cfg)      # inspect res.black / res.silent / res.cuts
run("input.mp4", "out.mp4", cfg)     # do the trim

The interval logic is exposed directly (merge_segments, intersect_segments, black_silent_cuts, keep_segments_from_cuts) and is pure/dependency-free.

Tests

pytest -q                  # all tests (integration generates a real clip)
pytest -q -m "not integration"   # pure unit tests only, no ffmpeg/OpenCV needed

The unit tests cover the interval algebra, the black-frame pixel classifier (including the stray-pixel tolerance), the silencedetect log parser, and the ffmpeg filtergraph builder. The integration test synthesizes a blue → black/silent → green clip with ffmpeg, runs the whole pipeline, and asserts the black stretch is detected and removed.

License

MIT — see LICENSE.