wjs-reframing-video

.crop.json

When to use

• Repurposing a 16:9 podcast / interview / talk for vertical short-video platforms (WeChat Channels 视频号, Douyin 抖音, Xiaohongshu 小红书, YouTube Shorts, TikTok, Reels).
• Repurposing a 9:16 phone recording for horizontal players (YouTube long-form, blog embeds).
• Repurposing 4:3 archive footage for 3:4 mobile, or vice versa.

When NOT to use

• Multi-person Q&A where each face needs its own crop — this skill picks one crop track per video. For per-speaker split renders, use wjs-editing-multicam instead.
• Animated content / B-roll with no faces — falls back to center crop, usually wrong for the intent.
• Heavy camera motion in the source (handheld pan/zoom) — the face tracker amplifies camera shake. Stabilize first.
• Source already at target aspect — no work to do.

What this skill IS — and IS NOT

--face-pick speaker

--face-pick largest

--motion cut

--motion smooth

FaceLandmarker

ffmpeg crop + scale

Dependencies

pip install mediapipe opencv-python numpy

PATH

face_landmarker.task

crop.py

~/.claude/skills/wjs-reframing-video/models/

Crop math

W / H

H / W

W_crop = H × H / W

H_crop = H

W_crop = W

H_crop = W × W / H

W_crop = 608

H_crop = 1080

W_crop = 1080

H_crop = 608

--output-size 1080x1920

Pipeline

1. Probe input dimensions, fps, duration via ffprobe.
2. Decide orientation — auto from aspect (

   --target portrait|landscape

   to override).
3. Sample frames at

   --sample-fps

   (default 5; high enough to catch mouth motion — Nyquist for speech is ~10 Hz, we need at least 4–5 fps).
4. Detect face landmarks per sampled frame with MediaPipe Tasks

   FaceLandmarker

   (478 landmarks). For each detected face record: center, size proxy, MAR (mouth-aspect-ratio = inner-lip vertical distance / horizontal mouth-corner distance).
5. Track faces across frames by center-distance matching → each face gets a stable

   face_id

   .
6. Per-sample active speaker: for each face track, variance of MAR over a sliding window (

   --mar-var-window-sec

   , default 1 s). The face with the highest variance is "speaking". Below

   --mar-var-threshold

   , no one is speaking → fall back to largest face.
7. Hysteresis: a candidate switch only commits if the new speaker is stable for

   --min-segment-sec

   (default 1.5 s). Shorter flickers are squashed — prevents the crop from ping-ponging on a one-frame mis-detection.
8. Speaker-aligned segments → for each segment, mean (cx, cy) of that speaker's face over the segment becomes the crop center, fixed for the full duration of the segment.
9. Build a ffmpeg step-function expression (

   --motion cut

   , default) that holds each segment's crop position constant and jumps instantly at each segment boundary — the visual feel of a real cut between camera angles. (

   --motion smooth

   switches to piecewise-linear pan between segment midpoints; rarely the right call for talking-head content because the camera appears to drift mid-sentence.)
10. Render one ffmpeg pass —

    crop=W:H:x='expr':y='expr', scale=OUT_W:OUT_H

    . The crop filter evaluates

    x

    and

    y

    per frame natively. Audio stream-copied.

scripts/crop.py

• <input>.crop.json

  — sidecar with the crop plan

• <input>_cropped.mp4

  — final cropped + scaled video

Sidecar schema (

<input>.crop.json

)

{
  "_about": "wjs-reframing-video crop plan for cam_a.MOV. Active-speaker detected via MAR variance.",
  "_help": {
    "source_size":     "[width, height] in pixels.",
    "target_size":     "[width, height] of the final rendered output.",
    "crop_window":     "[width, height] of the moving crop in source coords.",
    "chunks":          "Speaker-aligned segments: {t0, t1, cx, cy, speaker_id}.",
    "face_pick_mode":  "speaker = MAR-variance active-speaker; largest = old behavior.",
    "speaker_id":      "Stable face track id. null means no face / silence fallback."
  },
  "schema_version": 2,
  "source": "cam_a.MOV",
  "source_size": [1920, 1080],
  "target": "portrait",
  "target_size": [1080, 1920],
  "crop_window": [608, 1080],
  "face_pick_mode": "speaker",
  "sample_fps": 5.0,
  "mar_var_window_sec": 1.0,
  "mar_var_threshold": 1.5e-4,
  "min_segment_sec": 1.5,
  "chunks": [
    {"t0":  0.0, "t1":  4.2, "cx": 808, "cy": 540, "speaker_id": 0},
    {"t0":  4.2, "t1": 11.6, "cx": 1182, "cy": 540, "speaker_id": 1},
    {"t0": 11.6, "t1": 14.0, "cx": 808, "cy": 540, "speaker_id": 0}
  ],
  "face_sample_count": 1234,
  "track_count": 2
}

Performance

• Detection is the slow step. On Apple Silicon at 2 fps sampling, expect ~10–20× realtime (a 30-min source detects in ~1–2 min). Bumping

  --sample-fps

  makes detection slower but tracking more responsive.
• Render is fast — single ffmpeg pass with hardware encode (

  hevc_videotoolbox

  on macOS). Often <1× realtime for a 1080p source.
• For very long sources (>200 chunks), the ffmpeg expression gets cumbersome; the script auto-downsamples chunk midpoints to keep the expression under ~200 control points.

Common pitfalls

• Mouth gestures aren't speech — a yawn, laugh, eating, or sucking-in-air all raise MAR variance. The detector can briefly mistake these for talking. For high-stakes content, eyeball the speaker timeline in the sidecar (the script prints a

  face#N: Xs on screen (Y%)

  summary) and re-run with a different

  --mar-var-threshold

  if needed.
• Side-profile or down-tilted faces — when a face is rotated >60° from camera, MediaPipe may fail to land mouth landmarks reliably, so MAR variance flatlines. The speaker fallback to "largest face" kicks in. If you have a long stretch of profile shots, consider

  --face-pick largest

  .
• Two faces with overlapping speech (interruption / talking over) — both faces have MAR variance, only one wins. The losing face is treated as listener. For accurate per-speaker tracking under crosstalk, use wjs-editing-multicam with separate cams.
• Long stretches of silence (B-roll, music) — falls back to largest face. If the largest face is wrong (e.g. a listener stays still while the speaker's mic feeds music), you'll see drift. Pre-segment around music-only sections.
• Source has burned-in lower-thirds / subtitles — for H→V, the lower band gets cropped out; for V→H, it stays but gets stretched. Strip burn-ins before running.
• Wide-angle / fish-eye lenses — landmarks miss faces near edges. Pre-correct distortion with

  ffmpeg lenscorrection

  first.
• Upscaling artifacts —

  608×1080 → 1080×1920

  is a 1.78× upscale and visible on sharp text. Render at native crop dims (

  --output-size 608x1080

  ) and let the platform upscale, if you have overlays you want to keep sharp.
• Output bitrate > platform limit — default is

  --bitrate 12M

  . WeChat Channels (视频号) caps at 10 Mbps; pass

  --bitrate 8M

  for that target.