Media Scenedetect

Media Scenedetect

Context: $ARGUMENTS

PySceneDetect is the go-to tool for reliable scene-change detection. It is superior to ffmpeg's built-in scdet filter, especially on anime, stylized, and low-contrast content. The scenedetect CLI chains a detector (detect-content, detect-adaptive, detect-threshold) with one or more commands (list-scenes, save-images, split-video, export-html) in a single invocation.

Quick start

• Detect scenes + dump a CSV: → Step 3 (scenedetect -i in.mp4 detect-content list-scenes)
• Split video at every cut: → Step 4a (split-video -m for stream-copy via mkvmerge)
• Generate thumbnails per scene: → Step 4b (save-images -n 3 -o thumbs)
• Auto-chapter a long file: → Step 4c (use helper script chapters subcommand)

When to use

• You want reliable shot boundaries for cut sheets, editorial, or auto-chapters.
• ffmpeg scdet is returning noisy / missed cuts on stylized, anime, or dark content.
• You need thumbnails per scene for a dashboard, chapter art, or video summary.
• You're building a pipeline that re-muxes segments into concat lists.

Step 1 — Install

pip install scenedetect[opencv]            # recommended (bundled OpenCV wheel)
pip install scenedetect[opencv-headless]   # headless servers, no GUI deps
pip install scenedetect                     # core only — OpenCV already present

Verify: scenedetect version. Splitting requires ffmpeg (default) or mkvmerge on PATH.

Step 2 — Pick detector + threshold

Detectors are mutually exclusive positional commands:

Threshold intuition: higher threshold ⇒ fewer scenes (more tolerant of change). Lower for dark or low-contrast content; raise for noisy / shaky handheld.

Minimum scene length: pass -m 2.0 (seconds) to the detector to drop micro-cuts under 2s.

Luma-only fast mode: detect-content -l skips HSV, uses only luma — cheap on CPU.

Step 3 — Run

Single-detector + list:

scenedetect -i in.mp4 detect-content list-scenes -o scenes.csv

Chain multiple outputs in one pass (detector runs once):

scenedetect -i in.mp4 \
  detect-content -t 27 -m 2.0 \
  list-scenes -o scenes.csv \
  save-images -n 3 -o thumbs \
  split-video -o parts -m

Global options worth knowing:

• -o DIR — global output directory (applies to all commands unless overridden).
• -s STATS.csv — write per-frame metric CSV (very useful for threshold tuning).
• --downscale N — process every Nth pixel (N=4 for 4K); massive speed-up.
• --start T, --end T, --duration T — limit analysis range (skip intros).
• -v quiet|error|warning|info|debug — verbosity.

Step 4 — Use the scenes

4a. Split the video

# Re-encode via ffmpeg — exact cuts, slow
scenedetect -i in.mp4 detect-content split-video -o parts

# Stream-copy via mkvmerge — fast, cuts only at keyframes (not frame-exact)
scenedetect -i in.mp4 detect-content split-video -m -o parts

split-video requires ffmpeg unless -m is passed (then needs mkvmerge). Output filenames default to $VIDEO_NAME-Scene-001.mp4 etc.

4b. Thumbnails per scene

scenedetect -i in.mp4 detect-content save-images -n 3 -o thumbs

-n N = images per scene (default 3: first / middle / last). JPG by default; add --jpeg-quality 95 or switch with --png. Useful for chapter art or contact sheet.

4c. Chapters

No first-class chapter exporter — convert list-scenes CSV to ffmetadata format (the helper script's chapters subcommand does this). Then:

ffmpeg -i in.mp4 -i chapters.txt -map_metadata 1 -codec copy out.mp4

4d. HTML report

scenedetect -i in.mp4 detect-content list-scenes save-images export-html -o report.html

Self-contained HTML with thumbnails + timecodes. Great for editorial review.

Available scripts

• scripts/scenedetect.py — stdlib wrapper with check, detect, split, thumbnails, html-report, chapters subcommands. All support --dry-run and --verbose.

Workflow

# Verify install
uv run ${CLAUDE_SKILL_DIR}/scripts/scenedetect.py check

# Detect, get scene list as JSON
uv run ${CLAUDE_SKILL_DIR}/scripts/scenedetect.py detect \
  --input in.mp4 --method content --threshold 27 --min-duration 2.0

# Split with stream-copy
uv run ${CLAUDE_SKILL_DIR}/scripts/scenedetect.py split \
  --input in.mp4 --outdir parts --stream-copy

# Per-scene thumbnails
uv run ${CLAUDE_SKILL_DIR}/scripts/scenedetect.py thumbnails \
  --input in.mp4 --outdir thumbs --per-scene 3

# HTML report
uv run ${CLAUDE_SKILL_DIR}/scripts/scenedetect.py html-report \
  --input in.mp4 --output report.html

# Build ffmetadata chapter file
uv run ${CLAUDE_SKILL_DIR}/scripts/scenedetect.py chapters \
  --input in.mp4 --output chapters.txt

Reference docs

• Read references/scenedetect.md for detector comparison, threshold tuning per content type, CSV/JSON/HTML formats, scdet vs PySceneDetect, and recipe book (auto-chapter DVD, commercial detection, video summaries).

Gotchas

• Detector then command chain. scenedetect -i FILE <detect-X> <cmd1> <cmd2> ... — detect-X runs first; every command after re-uses its output. ALL can run in one invocation.
• Threshold scale is content-dependent. 15–40 typical for detect-content; default 27. Higher = fewer scenes. Lower for dark / low-contrast; raise for noisy / shaky.
• split-video defaults to re-encoding via ffmpeg — slow. Pass -m to stream-copy with mkvmerge — fast, but cuts fall on keyframes only (not frame-exact).
• save-images default is 3 per scene (first / middle / last). -n N overrides.
• -m on the detector sets minimum scene duration (seconds). -m on split-video means "use mkvmerge". Same flag letter, different meaning per command.
• HSV color space used internally for detect-content; on grayscale content use -l (luma-only) for speed.
• Default detects every transition, including fades. For mixed pacing use detect-adaptive.
• Large files: add --downscale 4 (process every 4th pixel). Cuts runtime massively with negligible accuracy loss.
• Always test threshold on a short clip first (--start 60 --duration 120) — threshold varies per content type.
• PySceneDetect is superior to ffmpeg scdet on anime / stylized / low-contrast content. Use scdet only when you already have ffmpeg in the pipeline and content is live-action.
• CSV header: Scene Number, Start Frame, Start Time (seconds), Start Timecode, End Frame, End Time (seconds), End Timecode, Length (frames), Length (seconds), Length (timecode).
• For frame-exact stream-copy cuts, convert the CSV to an ffmpeg concat list and re-mux with -c copy — but accept that cuts align to the nearest keyframe anyway.
• No ffmpeg required unless splitting. Analysis itself only needs OpenCV / PySceneDetect.
• --start + --duration skip intros / limit range; applies globally.

Examples

Example 1: Auto-chapter a DVD rip

scenedetect -i movie.mkv \
  detect-adaptive -m 5.0 \
  list-scenes -o scenes.csv save-images -n 1 -o chapter_art
# then convert scenes.csv → chapters.txt via helper script
uv run scripts/scenedetect.py chapters --input movie.mkv --output chapters.txt
ffmpeg -i movie.mkv -i chapters.txt -map_metadata 1 -c copy movie_chap.mkv

Example 2: Detect commercials in a TV recording

# Fades to black are the commercial breaks — use threshold detector
scenedetect -i recording.ts detect-threshold -t 8 list-scenes -o breaks.csv

Example 3: Video summary grid

scenedetect -i vlog.mp4 detect-content -t 30 -m 3.0 save-images -n 1 -o frames
# then montage via ImageMagick
magick montage frames/*.jpg -tile 6x -geometry 320x180+4+4 summary.jpg

Example 4: Fast first pass on a 4K file

scenedetect --downscale 4 -i 4k.mp4 detect-content -l -m 2.0 list-scenes -o scenes.csv

Troubleshooting

Error: scenedetect: command not found

Cause: not installed or not on PATH. Solution: pip install scenedetect[opencv]. If installed in a venv, activate it or use python -m scenedetect ....

Error: Failed to open video: <file> / cv2.error

Cause: OpenCV can't decode the container (some MKVs with odd codecs). Solution: re-mux to mp4 first (ffmpeg -i in.mkv -c copy in.mp4) or install full scenedetect[opencv] wheel.

Too many / too few scenes

Cause: threshold off for your content. Solution: generate a stats file (-s stats.csv) to see per-frame metric, then pick threshold above typical baseline but below scene-cut spikes.

split-video produced weird-length segments

Cause: -m (mkvmerge) only cuts on keyframes — GOP-aligned, not frame-exact. Solution: drop -m to re-encode (frame-exact) or raise GOP density at encode time.

Adaptive detector misses obvious cuts

Cause: rolling window smoothing threshold too high for slow-paced content. Solution: lower -t (e.g. -t 2.0), or fall back to detect-content with a high min-duration.