/klik-cam

Camera ← music recipes — implementing CAM-5 / CAM-6

The preset family (field-judged)

Named single moves (truck/dolly/dolly-zoom/whip/rack focus…) live in references/camera-moves.md — preset A composes its glides from that vocabulary.

Continuity trap (render-found): every always-on layer of the pose — breathing, drift, swell — must be keyed to the global frame, never to the section index. A section-keyed phase (sin(f/95 + section·1.7)) teleports its offset the frame the index increments: a one-frame pop at every glide landing. Verify seams numerically: frame-difference the render across each boundary — the landing frame's diff must be in line with its neighbors. / CAM-6

Load this when a piece's camera should follow the music. It turns the method's analyzer output (music map + events + levels) into camera movement. Every recipe was render-verified in a 2.5D POC against a real track. Two presets, field-judged:

PRESET A — HOLD & GLIDE (the default for motion graphics)

The camera holds on each section's pose — Anchored-level breathing only (≤0.5% creep) — then executes ONE composed zoom/slide transition per musical seam. Field verdict drove this: continuous reactive drift "reads as camera shake, not motion graphics."

• Scenes live side by side in world space (one slot per music-map section, e.g. 2400px apart); the camera travel IS the scene change.
• Arrive on the seam: the glide STARTS T frames before the next section's boundary and LANDS exactly on its downbeat — the move belongs to the music. T = clamp(56 − 6·tier_incoming, 24, 50): hot section ⇒ snappy arrival, calm ⇒ long glide. Monotone ease-in-out — no overshoot (C0 rule).
• Poses are quantized per section (zoom level ∈ {0, 80, 150} · y ∈ {−45, 0, 45} · roll ∈ {−0.9°, 0, 0.9°}, seeded) so each transition reads as a deliberate zoom-in / zoom-out / rise / drop — CAM-1 variation without wander.
• Holds carry only breathing + a gentle level swell on depth — during element phrases the camera is still by construction (CAM-6 satisfied structurally: it moves BETWEEN element phrases, at the seams).

PRESET B — REACTIVE SHAKE (opt-in accent flavor, never the default)

The sidechained continuous rig below (drift + bar-kick hits). Field-judged too twitchy as a standing behavior — use it scoped: a hype section, a performance montage, a deliberate handheld-energy moment. The user opts in; record the scope in the spec.

ROADMAP — ELEMENT-FOLLOW (next)

The third preset: the camera framing driven by ELEMENT positions (follow a travelling focal with inertia, reframe on staging changes, gaze-anchored homes per A12) with music as the timing layer. Deliberately deferred.

Preset B mechanics (the original CAM-5 rig)

Element-side mappings live in klik-anim references/music-motion.md — the camera takes the music's big body (bars, sections, drops); elements take the syncopation and the surface detail. Do not mix the layers.

The rig shape (CAM-3 compliant)

outer AbsoluteFill        perspective: 1100        ← NO filters, ever
└─ THE SPACE  div         preserve-3d; the camera is EXACTLY ONE transform here:
   │                      translate3d(−x,−y,z) rotateZ(−rz) rotateY(ry)
   ├─ far plane   div     translateZ(−380) · DOF blur 2px   (parked, soft)
   ├─ mid plane   div     translateZ(0)    · crisp          (FOCAL — CAM-2)
   └─ near plane  div     translateZ(240)  · DOF blur 1px   (biggest parallax)

Lateral camera translate + perspective produces true parallax (near planes shift more on screen); pair it with a small rotateY (ry = x * 0.008) so no move is ever a pure truck (anti-AI-camera).

The verified mappings

DISCOVERED TRAP — sample-stacking motion blur flattens 3D

@remotion/motion-blur's CameraMotionBlur composites fractional-opacity samples. Opacity is a CSS grouping property: it forces transform-style: flat, so a preserve-3d camera space inside it collapses — mid/near planes vanish (render-verified failure; the far plane's leftovers made it look like a partial render). This is CAM-3's ancestor-filter law in another costume: no filter, no opacity < 1, no grouping property on the space or any ancestor of it.

• DOM 3D rig → implement the smear as the per-plane velocity blur above.
• CameraMotionBlur remains correct for flat scenes (camera faked in 2D transforms without preserve-3d) — that's its domain, not this one's.

Core code (verified — src/music-cam.ts)

export const barKicks = kicks.filter((e) => {
  const pos = ((e - DOWN0) % BAR + BAR) % BAR;
  return pos <= 2 || pos >= BAR - 2;                 // the big body only
});

const signature = (i: number) => ({                   // CAM-1, per section
  ax: random(`sig-ax-${i}`) * 2 - 1,
  ay: (random(`sig-ay-${i}`) * 2 - 1) * 0.55,
  rz: (random(`sig-rz-${i}`) * 2 - 1) * 0.9,
  period: 110 + random(`sig-p-${i}`) * 130,
  phase: random(`sig-ph-${i}`) * Math.PI * 2,
  mag: 34 + random(`sig-m-${i}`) * 46,
});

// camPose(frame): section drift (crossfaded 36f at seams, smoothstep)
//   + swell push from smoothed levels + bar-kick hit (z/rz/x paired, decay 5)
//   + ry = x * 0.008  — full implementation in the POC, structure above.

Preset C — ELEMENT-FOLLOW (render-verified)

CAM-6 doctrine applied to pursuit. Three mechanisms, all pure functions of frame:

• The landing track. The follow target is where the focal will LAND, not where it is: the moment a hop/phrase starts, the target jumps to its end point. The element's arcs, squash, and bounces play out INSIDE a stable frame — the camera tracks phrases, the element keeps the syncopation.
• Lookahead lead room. Sample the landing track ~14f into the future (deterministic — the future is data) to counter the inertia lag: the subject then rides slightly behind center with clear space ahead — real follow grammar. Without it the camera chases (subject ahead of center — a render-found defect, fixed and re-verified).
• Inertia. Camera x = trailing exponential average of the target (Σ target(f−k)·e^(−k/10), k ≤ 36) — steadicam glide, natural lag on direction changes, settle-into-breath. Breathing/swell layers key to the GLOBAL frame (continuity law).

Boss handoff (A12 eye-trace made physical): when the focal changes, the target blends to the new boss over ~30f with smoothstep, keyed to the musical seam — never a snap. And the outgoing boss must actually YIELD: render-found, the traveller kept advancing on sparse breakdown kicks and wandered past its replacement mid-handoff — a resting boss's residual event stream is texture, not travel; freeze its phrase motion for the handoff span.

THE GOVERNOR — the jarring threshold. Too much camera movement back-to-back is jarring regardless of what demands it. Two caps, both render-found and numerically verified:

• Element phrases complete. An event may only start a travel phrase if the previous phrase has landed (greedy min-gap = phrase duration + a ground beat). Without it, dense syncopation (kicks 2–5f apart vs a 12f hop) snaps the base position forward — the subject teleports. Events beyond phrase capacity are texture, not travel.
• Sustained camera speed cap (~13px/f at 1080p): a forward pass clamps the follow target's per-frame advance, so demand bursts become a steady glide toward the trend instead of surge-after-surge — the operator going steady-and-wide when the subject gets frantic. Verify numerically: max per-frame camera delta ≤ cap across the piece; the focal's max delta ≤ its own phrase cruise speed.

Music layer on top: tier can scale follow tightness; depth-only co-fire on bar-kicks stays tiny — the travelling element owns its own impacts — and is gated OFF entirely while the boss is not the traveller (render-found: a bar-aligned kick during the handoff pushed the camera while it contemplated a resting boss — jitter on a moment that must be still).

Per-segment physics (element side, feeds the follow): when travel energy evolves with the piece, keep gravity constant and derive airtime from bounce — T = √(8h/g); height and stride scale with the section tier, impact squash/ripple scale with height (A16 weight). Size the governor cap just above the fastest segment's cruise speed (max(stride/T)) so steady travel passes while bursts still cap.

CAM-6 layer — the sidechain (v2, render-verified A/B)

v1 of this rig moved the camera off the music alone and walked the subject off-center during busy element phrases. v2 subordinates it (the DEFAULT; user-overridable as a recorded spec decision):

// authored element motion — the SAME music data, used a second time:
elementMotion(f) = min(1, kickPulse + 0.5·snarePulse + 0.2·hatPulse
                        + |Δ bass level|·8 + |Δ vocal level|·5)

// sidechain envelope — fast attack (current value counts fully),
// slow release (~24–30f perceptible tail):
duckEnv(f) = max over k∈0..30 of elementMotion(f − k) · exp(−k / 12)

// the camera's claim on the residual:
DIAL_GAIN = [—, 0, 0.18, 0.5, 1.0, 1.5]        // C0 dial = claim multiplier
gXY = tierScale(tier) · DIAL_GAIN[dial] · (1 − 0.85·duck)   // frame plane
gZR = tierScale(tier) · DIAL_GAIN[dial] · (1 − 0.45·duck)   // depth/roll
// → complementary axes: elements own the frame plane, camera's residual
//   lives in depth. Margin cap AFTER gains: |x| ≤ 55px, |y| ≤ 40px (A11
//   inner pocket at 1080p). Accent co-fire (bar-kicks, tier ≥ 4) is EXEMPT
//   from the duck, scaled by the dial, and rides z/rz with only ~4px x.

Verified A/B on the same track and frames: busy tier-5 frame — v1 subject displaced right + tilted, v2 dead-center with camera parked; bar-kick frame — v2 keeps the push + smear co-fire with composition held. The music is still audible in the camera — it just speaks in the gaps the elements leave.

Judging it (CAM checklist hooks)

• Two stills in different sections show different drift orientations (CAM-1).
• A breakdown still is crisp and near-home; a hit still is displaced AND smeared; 3 frames later it is sharp again (checklist 3–4).
• The focal plane stays crisp through every drift (CAM-2) — whole-frame smear is allowed only at the 2–4 frame velocity peak.
• Scrub any section seam: the signature changes, the camera never jumps.
• CAM-6: pick the busiest element frame — the focal centroid sits inside the inner margin and the camera is near-parked; pick a designed hold — the camera carries the life. If both fail, the sidechain isn't wired.