SafeZoneDetectorV2

Replaces the V9 heuristic detector. Goal: never crop the facecam, crosshair, HUD, minimap, or killfeed when reframing gameplay to 9:16.

V9 defects this fixes

Detected zones are discarded. safe-zone-detector.js:238 builds a zones array, then returns new SafeZoneMap(...) without calling addZone. So every detection is silently thrown away and safeZones is always empty. V2 must add zones to the returned map.
Single synthetic frame. V9 detection runs on one frameData blob with no real multi-frame sampling. Facecams and killfeeds are defined by change over time; one frame can't see that.
No crop-priority enforcement. V9 lists conflicts but doesn't rank what to protect first.

Inputs

V2 samples N frames across the clip via the existing ffmpeg raw-frame pipeline (same approach as highlight-engine.detectMotion):


ffmpeg -i <video> -vf fps=<s>,scale=<w>:<h> -f rawvideo -pix_fmt rgb24 -

Default: sample everys, downscaled (e.g. 320×180) for speed. Per-region temporal variance distinguishes static UI (HUD) from changing content (killfeed, facecam).

Detected region types

Type	Typical location	Signal
`facecam`	a corner, 15–25%	high temporal variance + skin-tone fraction + rectangular/round edge
`hud_top`	top band	persistent high-contrast text/icons, low motion
`hud_bottom`	bottom band	persistent bars/numbers (health/ammo)
`minimap`	a corner, 10–15%	dense edges, colored, semi-static
`killfeed`	top-right	text rows that change rapidly
`crosshair`	center, tiny	persistent center mark
`subtitles`	lower third	existing burned captions

Protected region output


{

  "videoWidth": 1920,

  "videoHeight": 1080,

  "regions": [

    { "type": "facecam", "bounds": {"x":0.75,"y":0.0,"width":0.25,"height":0.25},

      "confidence": 0.0, "framesSeen": 0, "priority": 1 }

  ],

  "cropPlan": { "format": "9:16", "x": 0.0, "width": 1.0, "preserved": ["facecam","crosshair"] },

  "detectedAt": "ISO-8601"

}

confidence is the fraction of sampled frames the region fired in — a measured quantity, never a constant. framesSeen exposes the denominator so the value is auditable.

Crop priority

When a 9:16 window cannot contain everything, protect in this strict order:


1. facecam        (creator identity — losing it kills the clip)

2. crosshair      (gameplay readability)

3. hud_*          (context: health/ammo/score)

4. gameplay action (the rest)

The cropper picks the 9:16 window that preserves the highest-priority set possible. If facecam + action can't both fit, it emits a picture-in-picture plan (facecam composited as an overlay) rather than cropping the facecam out.

Confidence & honesty

A region with confidence < 0.5 is reported as candidate, not confirmed, and the UI labels it as such.
If detection can't run (ffmpeg missing / unreadable video), V2 returns { status: "unavailable", reason } — it does not emit a fabricated default layout.

Feature flag

Behind safeZoneV2 (LANTERN_CI_SAFEZONE_V2, default off). When off, the editor uses the V9 detector (with the discard bug fixed as a minimal patch) so reframing still functions.

Implementation steps

Minimal V9 patch: add detected zones to the returned SafeZoneMap (fixes the discard bug today).
Multi-frame sampler reusing the ffmpeg raw pipeline.
Per-region temporal-variance + skin-tone + edge-density scoring → confidence.
Crop planner with the priority ladder + PiP fallback.
Wire into routes/creator-entries.js render path behind the flag.