Wjs Editing Multicam

wjs-editing-multicam

Combine N synced camera angles into a single rendered MP4. Decisions are audio-energy-driven only — the cam with the loudest mic each second wins. Output is hard cuts (or hard cuts plus a corner PiP).

Setup & commands

The implementation lives in the open-source polysync pip package (https://pypi.org/project/polysync/ · https://github.com/jianshuo/polysync) — this skill no longer ships its own scripts. Install it, then drive it via its CLI:

python3 -m pip install -U polysync      # needs ffmpeg/ffprobe on PATH

polysync edit        CAM_A CAM_B CAM_C --out edl.json   # build the decision list
polysync render-cuts edl.json --out out.mp4             # hard cuts
polysync render-pip  edl.json --out out.mp4 --pip bottom-right   # cuts + corner inset

edit and the renderers read each input's .sync.json automatically. Sync first with wjs-syncing-multicam (polysync sync) if the sidecars don't exist yet.

Render flags for raw camera footage (see Preflight below for when to use each):

# Sony S-Log3 footage, shot vertical, FX6 cams turned on their side, for 小红书:
polysync render-cuts edl.json --out out.mp4 \
    --log slog3 \           # S-Log3/S-Gamut3.Cine -> Rec.709 grade
    --rotate 1:90 --rotate 2:90 \   # rotate cam1,cam2 90° CW (FX6 with no flag)
    --width 1080 --height 1920 --fill \   # vertical, crop-to-fill (no black bars)
    --duck-audio --audio-cams 0,1        # clean speaker-gated audio (cams 0,1 = the two lavs)

--duck-audio replaces the single-cam soundtrack with a speaker-gated mix: each moment keeps the active speaker's close mic and ducks the rest (much cleaner than a constant 2-mic sum, which piles up bleed/room tone). --audio-cams 0,1 restricts gating to the real speaker mics — always exclude the wide/room cam, whose mic sits at a similar level but is reverby and would otherwise get picked. (See the editing-quality note below for the why.)

Preflight — ALWAYS check raw footage before rendering (hard-won)

Straight-off-the-card footage renders WRONG without these checks. Spend 2 minutes here or you'll render the whole thing broken (we did).

1. Color profile (Log). Sony FX3/FX6 default to S-Log3 / S-Gamut3.Cine — flat, grey, low-contrast. It MUST be converted to Rec.709 or it looks washed-out and broken. Check the .XML sidecar's CaptureGammaEquation (s-log3-cine) or ffprobe -show_entries stream=color_transfer. Fix: --log slog3. (Performance: the LUT is applied AFTER downscale automatically — a 3D LUT on 4K is ~4x slower than on 1080p for an identical result.)
2. Orientation. Phone / vertically-mounted shoots record rotated. Extract one frame per cam (ffmpeg -ss 200 -i CAM -frames:v 1 f.jpg) and LOOK. Some cameras (FX3) write a rotation flag → ffmpeg auto-rotates, fine. Others (FX6 physically turned on its side) write no flag → people come out lying down. Fix per-cam: --rotate 1:90 (90 = clockwise; try 270 if upside-from-the-other-side).
3. Delivery orientation. 小红书 / Reels / Shorts are vertical — and this footage is usually shot vertical. Render --width 1080 --height 1920 --fill. Default (1920×1080, pad) is for landscape only; mixing a portrait source into a landscape frame pillarboxes it (ugly black side bars).
4. Staggered camera starts. Cameras rarely roll at the same instant; the sidecar delta/coverage shows it. The opening N seconds may be single-cam (only the first-rolling camera covers t=0). That's unavoidable — expect a single-cam intro of max(delta) seconds.

Editing quality — when polysync edit isn't enough

polysync edit switches to the loudest mic per second. With close, bleeding mics (each cam's mic also picks up the other speaker loudly), the close/guest mic stays loudest even when the other person talks, so the editor over-selects that cam and the cuts don't track the real speaker. Two fixes, applied by hand on the EDL when "cut to the correct speaker" matters:

• Per-mic baseline-normalized speaker attribution. Per second, subtract each mic's own median energy, then pick whichever mic is highest relative to its own baseline → that's who's actually talking. (polysync's raw cam[k] - mean(others) doesn't remove the per-mic baseline offset, so it favors the loud mic.)
• Cutaways for rhythm (剪辑感). Audio attribution alone gives long static holds. Cap any single shot (~8–10 s) and insert ~3 s cutaways to the listener (reaction shot) and the wide cam, alternating. The wide cam's quiet mic means the editor never picks it on energy — inject it deliberately for establishing / 整体.

What this skill IS — and IS NOT

If you need face tracking, fade transitions, captions, or HyperFrames composition, use the hyperframes skill on top of this skill's MP4 output.

REQUIRED INPUT

Original camera files (untouched) plus their .sync.json sidecars next to them. If sources aren't synced yet, run wjs-syncing-multicam first to write the sidecars. Missing sidecar = cam assumed at delta=0, full coverage.

polysync edit reads each sidecar for delta_seconds + overlap_in_reference, lifts the cam's audio envelope into the reference timeline, and only schedules a cam during its coverage window. polysync render-cuts / polysync render-pip apply ffmpeg -itsoffset per input using the EDL's deltas[] array.

When NOT to use

• One source — nothing to switch between; use video-segmentation.
• Polished NLE timeline already exists — don't fight the editor.
• Want fade transitions / overlay captions / brand title cards — run this skill first to get the cut-down MP4, then feed it into wjs-overlaying-video or hyperframes.

Pipeline

1. Read each input's sidecar → list of delta_seconds[k] + overlap_in_reference[k].
2. Extract per-cam mono PCM @ 16 kHz from the original file.
3. Log-RMS envelope at 1 Hz frame rate (per-second).
4. Lift each envelope into reference timeline by indexing at t_ref - delta_k; uncovered seconds become -inf so they're never picked.
5. Audio source = the cam with the largest envelope spread (90th − 10th percentile over its covered seconds), with a small bonus for coverage fraction.
6. Score per second: cam[k] - mean(other covered cams). Highest score = best active-speaker candidate.
7. Editor decides EDL — two modes:
   • rotation (default): random dwell in [min_dwell=8, max_dwell=15] s, pick best-scoring covered cam (≠ current) at each cut.
   • greedy: hysteresis — hold current unless another cam's lookahead-window score beats it by --switch-threshold. Floor min_dwell=4, ceiling max_dwell=18. Both force-switch if the active cam exits its coverage window mid-shot. If no cam covers t=0 (overlap windows that start a few seconds in), the editor opens at the first covered second and backfills the lead-in with cam 0.
8. Emit EDL JSON.

EDL schema (edl.json)

{
  "_about": "EDL produced by polysync.edit.autoedit. Times in reference timeline. Render commands apply ffmpeg -itsoffset deltas[k] per input.",
  "_help": {
    "inputs":        "Original media paths, in cam-index order (cam 0, cam 1, ...).",
    "deltas":        "Per-cam delta_seconds from each sidecar. Render uses ffmpeg -itsoffset deltas[k].",
    "duration_sec":  "Output duration in reference timeline.",
    "audio_source":  "Cam index whose audio track becomes the master. Single source — not a mix.",
    "coverage":      "[start, end] per cam in reference timeline.",
    "edl":           "List of {cam, start, end} segments. Times are reference-timeline seconds."
  },
  "inputs":       ["cam_a.MOV", "cam_b.MOV"],
  "deltas":       [0.0, 12.345],
  "duration_sec": 4512,
  "audio_source": 0,
  "coverage":     [[0.0, 4512.0], [12.345, 4499.835]],
  "edl":          [{"cam": 0, "start": 0, "end": 13}, {"cam": 1, "start": 13, "end": 28}, ...]
}

polysync edit writes _about + _help directly into the file so opening the JSON in any editor explains itself.

Render

Both apply -itsoffset deltas[k] per input.

Brainstorm before running

Three real knobs to confirm with the user:

• Pacing — --mode rotation (varied dwell, easier on the ear) vs --mode greedy (energy-following, snappier).
• PiP — yes / no. If yes, which corner + width?
• Min cut length — --min-dwell floor. 8 s default for rotation is conservative; talking-heads can go to 4.

audio_source is auto-picked; override with --audio-source <cam-index> if the auto-pick sounds wrong on a 30 s listen.

File layout

working_dir/
  cam_a.MOV                 # ORIGINAL, untouched
  cam_a.MOV.sync.json       # from wjs-syncing-multicam
  cam_b.MOV                 # ORIGINAL, untouched
  cam_b.MOV.sync.json
  edl.json                  # from `polysync edit`
  multicam_render.mp4       # from `polysync render-cuts` OR `polysync render-pip`

Common pitfalls

• Trusting audio_source without listening. Spread + coverage is a proxy. Always sample a 30 s clip before committing — a high-spread track can still be clipped / distorted.
• Running polysync edit on the full 75 min before tuning. Run on a 2-min slice first (ffmpeg -ss A -t 120 an extract per cam), listen, adjust --min-dwell / --mode, then commit to full length.
• Expecting face-driven framing. This skill doesn't see the video — only the audio. If one cam is well-framed but quiet, the editor won't favor it. Use --audio-source + per-segment pip overrides as the manual escape hatch.
• Re-rendering when sync was wrong. EDL bakes in deltas[] at edit time. If you fix the sidecars later, re-run polysync edit to regenerate the EDL before re-rendering.
• Rendering a mid-video window straight from the long 4K originals = brutally slow. polysync render-cuts trims in the filter graph (trim=start=…) with no input -ss seek, so for a clip whose EDL starts deep in the source (e.g. a 90 s segment 28 min into a 34 min file) ffmpeg DECODES the 4K from t=0 to reach the window — ~20 min for a 92 s clip. Fix: pre-cut each cam's window first with accurate fast seek (ffmpeg -ss <start-2> -i src -ss 2 -t <dur> … — fast-seek to 2 s before, then accurate-decode 2 s), already rotated + graded + scaled to the target frame, then concat the multicam EDL on those short prepared windows (0-based times, cheap). All cams cut to the SAME reference window (source-local start = ref_start - delta_k) so they stay in sync. This is how /wjs-segmenting-video should render multicam segments — never feed a mid-video absolute-time EDL to a from-zero decode.