openclaw-video-editor（OpenClaw video editor 视频）

openclaw-video-editor

v4.2.0

A small, honest video editing skill for OpenClaw built on ffmpeg, ffprobe, and python3. It runs entirely on the local machine. No external services, no AI segmentation, no transcription model.

Scope

This skill provides:

• scripts/check_deps.sh — verify ffmpeg, ffprobe, python3 are installed before any workflow runs.
• scripts/generate_srt.py — convert Whisper / Deepgram / AssemblyAI / generic word-timing JSON into .srt, .vtt, or .ass. Bug-fixed: chunk text no longer leaks across segment boundaries.
• scripts/highlight_reel.py — detect scene changes via ffmpeg and assemble a short highlight reel.
• scripts/apply_lut.py — apply a real .cube 3D LUT (lut3d), or use a named filter preset (warm, cool, bw, high-contrast, faded).
• scripts/loudnorm_two_pass.py — automate the two-pass loudnorm workflow end-to-end (NEW in v4.2.0).
• scripts/make_vertical.py — letterbox or center-crop a horizontal video into a 9:16 vertical with optional blurred-background fill (NEW in v4.2.0).

This skill does not perform:

• AI background removal or subject segmentation
• Voice cloning or generative video
• Transcription (it only formats word timings into subtitle files; pair with a separate STT tool)
• Any remote API calls

Required binaries

bash scripts/check_deps.sh

Returns non-zero if ffmpeg, ffprobe, or python3 is missing.

Workflows

1. Generate subtitles from a transcript

python3 scripts/generate_srt.py transcript.json subtitles.srt
python3 scripts/generate_srt.py transcript.json subtitles.vtt
python3 scripts/generate_srt.py transcript.json subtitles.ass --font Helvetica --fontsize 28

Tunable: --max-chars, --max-words, --max-duration for non-English language tuning.

2. Burn subtitles into a video

ffmpeg -i input.mp4 \
  -vf "subtitles=subtitles.srt:force_style='Alignment=2,MarginV=30,Outline=2,Shadow=1'" \
  -c:a copy output.mp4

Alignment=2 is bottom-center. Use Alignment=8 for top-center.

3. Background blur (privacy filter, not subject segmentation)

True shallow-depth-of-field requires AI segmentation, which this skill does not include. The filter below is a full-frame blur:

ffmpeg -i input.mp4 -vf "boxblur=20:1" -c:a copy blurred.mp4

If you already have a binary alpha matte (mask.mp4, subject white, background black, frame-aligned), you can composite a blurred background behind the original subject:

ffmpeg -i input.mp4 -i mask.mp4 \
  -filter_complex "[0:v]boxblur=20:1[bg];[0:v][1:v]alphamerge[fg];[bg][fg]overlay=format=auto" \
  -c:a copy composited.mp4

Producing the matte itself is out of scope.

4. Color grading

Filter presets (no LUT file required):

python3 scripts/apply_lut.py input.mp4 - graded.mp4 --preset warm
python3 scripts/apply_lut.py input.mp4 - graded.mp4 --preset bw

Real .cube LUT:

python3 scripts/apply_lut.py input.mp4 lut.cube graded.mp4 --strength 0.8

5. Audio normalization (one command, NEW in v4.2.0)

Previously the SKILL.md only described the two-pass loudnorm flow without automating it. v4.2.0 ships a runner that does both passes:

# Broadcast (-23 LUFS, EBU R128)
python3 scripts/loudnorm_two_pass.py input.mp4 normalized.mp4

# Streaming platforms (-14 LUFS)
python3 scripts/loudnorm_two_pass.py input.mp4 normalized.mp4 --target-lufs -14

The script runs Pass 1, parses the JSON measurement block from ffmpeg's stderr, and runs Pass 2 with the measured offsets applied. Video is copied without re-encoding.

6. Highlight reel via scene detection

python3 scripts/highlight_reel.py input.mp4 highlight.mp4 --duration 30 --threshold 0.4

7. Watermarking

# Bottom-right text watermark
ffmpeg -i input.mp4 \
  -vf "drawtext=text='@yourhandle':x=w-tw-20:y=h-th-20:fontsize=24:fontcolor=white@0.7:box=1:boxcolor=black@0.4:boxborderw=8" \
  -c:a copy watermarked.mp4

# Image overlay (logo)
ffmpeg -i input.mp4 -i logo.png \
  -filter_complex "[0:v][1:v]overlay=W-w-20:20" \
  -c:a copy watermarked.mp4

8. Make vertical 9:16 for shorts (NEW in v4.2.0)

make_vertical.py offers three modes:

# Letterbox: original aspect inside a black 1080x1920 frame
python3 scripts/make_vertical.py input.mp4 vertical.mp4 --mode letterbox

# Crop: center-crop to 9:16
python3 scripts/make_vertical.py input.mp4 vertical.mp4 --mode crop

# Blur-fill: original centered, blurred copy of itself fills the bars
python3 scripts/make_vertical.py input.mp4 vertical.mp4 --mode blur-fill

blur-fill is the most popular look for repurposing horizontal content as shorts.

9. Format conversions

# 1080p H.264 with reasonable quality
ffmpeg -i input.mp4 -vf "scale=-2:1080" \
  -c:v libx264 -preset medium -crf 20 \
  -c:a aac -b:a 160k output_1080p.mp4

Safety properties

• All workflows run locally. No remote API calls.
• Python helpers use subprocess.run with argument lists (never shell=True), and reject paths with shell metacharacters via a strict regex allowlist.
• The skill never modifies system configuration, environment variables, or other plugins.
• The skill only reads input files and writes output files at the paths the user provides.

Known limitations

• Inputs must be valid video/audio files reachable on the local filesystem.
• The blur compositing path requires a pre-existing per-frame alpha matte. Producing the matte is out of scope.
• make_vertical.py --mode blur-fill does a full re-encode; the other modes also re-encode the video stream. Audio is copied where possible.

v4.2.0 changes

• Description on the registry now matches the actual workflows. The previous "Pro-Studio AI background removal" copy contradicted the SKILL.md and triggered a registry mismatch warning even after the SKILL.md was rewritten in v4.1.0. v4.2.0 republishes the consistent description so the public listing and SKILL.md agree.
• scripts/loudnorm_two_pass.py automates the previously-manual broadcast loudness flow.
• scripts/make_vertical.py adds three vertical-conversion modes for shorts (letterbox, crop, blur-fill).
• Existing scripts (generate_srt.py, highlight_reel.py, apply_lut.py, check_deps.sh) are unchanged from v4.1.0; their bug fixes carry over.

License

MIT. See LICENSE.