CaptionEngineV3

Builds on caption-engine.js (V9: Caption class, style presets, highlight-driven generation). V3 adds an event taxonomy, dynamic caption text tied to detected events, a 50+ style library, and visual effects.

Event taxonomy

Captions are triggered by events derived from real signals the highlight engine already produces (motion spikes, audio spikes, scene cuts) plus optional speech transcript. No event is invented; each carries the timestamp and the signal that produced it.

Event	Derived from	Example trigger
`kill`	audio spike + scene cut	weapon/hit transient
`multi_kill`	≥2 `kill` events within window	clustered transients
`clutch`	sustained motion + late audio spike	end-of-round tension
`near_death`	motion dip + low-audio then spike	comeback moment
`jump_scare`	sharp scene cut + audio spike	sudden change
`funny`	speech/laughter (if transcript)	reaction
`victory`	end-region scene change + audio	win screen
`loss`	end-region scene change	defeat
`reaction`	facecam-region motion (SafeZoneV2)	streamer reacts

Each event:


{ "type": "clutch", "t": 12.4, "confidence": 0.0, "evidence": ["[email protected]","[email protected]"] }

confidence and evidence make every caption auditable — you can see why it fired. Events below a confidence floor are dropped rather than captioned on a guess.

Dynamic caption text

A template bank maps event → candidate phrases (the editor picks one, or the operator edits):


kill        → "GOT HIM" · "DOWN" · "ONE SHOT"

multi_kill  → "NO WAY" · "DOUBLE" · "HE'S CRACKED"

clutch      → "CLUTCH" · "WATCH THIS" · "WAIT FOR IT"

near_death  → "THAT WAS CLOSE" · "HOW?!"

victory     → "GG" · "THAT'S A WIN" · "THAT CHANGES EVERYTHING"

reaction    → "HE DID WHAT?" · "THIS WAS INSANE"

Text is a suggestion, always operator-editable. The engine never asserts a caption describes reality beyond the signal that triggered it.

Style library (50+)

Styles are data, not code — defined in a caption-styles.json so the count is real and extensible. Families:

mrbeast (bold, high-contrast, drop shadow)
gaming_montage (impact, shake, glow)
tiktok_gaming (rounded, bouncing)
reaction_creator (speech-bubble)
esports (clean, lower-third)
streamer_highlight (outlined, animated entry)

Each style:


{

  "id": "mrbeast_bold",

  "family": "mrbeast",

  "fontFamily": "Impact, sans-serif",

  "fontSize": 64,

  "color": "#ffffff",

  "outline": { "color": "#000000", "width": 4 },

  "shadow": { "color": "rgba(0,0,0,0.8)", "blur": 6 },

  "animation": "scale_pop",

  "safeMargin": 60

}

Visual effects

bounce · scale_pop · shake · impact_flash · glow · outline · drop_shadow — rendered via ffmpeg/libass (ASS override tags) or a frame compositor. Effects are presentational only; they never change the underlying event truth.

Mobile-safe placement

Captions respect SafeZoneV2 output: never placed over facecam/HUD/killfeed; default lower-third with a configurable safe margin (≥60 px at 1080×1920).

Outputs

.ass/.vtt subtitle track (for preview without re-encode)
burned-in render (for final export) — verified by ExportValidator ("captions actually rendered")

Feature flag

Behind captionEngineV3 (LANTERN_CI_CAPTION_V3, default off). When off, V9 caption generation remains.