Communitygithub.com

inkbeat-shorts

Toon-shaded 3D cartoon videos in Remotion + React Three Fiber, for vertical Shorts and calm 16:9 films. Covers beat-driven camera work, post-process ink lines, color grading, flicker hunting, deterministic multi-tab rendering, motion blur, captions, sound and QA, with a ready starter project. Use it to start or fix a 3D cartoon video in Remotion, or when a render flickers, shimmers, shows broken outlines or changes between runs.

inkbeat-shorts란 무엇인가요?

inkbeat-shorts is a Claude Code agent skill that toon-shaded 3D cartoon videos in Remotion + React Three Fiber, for vertical Shorts and calm 16:9 films. Covers beat-driven camera work, post-process ink lines, color grading, flicker hunting, deterministic multi-tab rendering, motion blur, captions, sound and QA, with a ready starter project. Use it to start or fix a 3D cartoon video in Remotion, or when a render flickers, shimmers, shows broken outlines or changes between runs.

지원 대상✓Claude Code~Codex CLI~Cursor
npx skills add https://github.com/k4878337-lang/inkbeat-shorts/tree/HEAD

즐겨 사용하는 AI에게 물어보기

이 에이전트 스킬이 미리 로드된 새 채팅을 엽니다.

문서

inkbeat-shorts: toon-shaded 3D Shorts and long-form videos in Remotion

A field guide distilled from building 3D cartoon Shorts and long-form films in Remotion + three.js / React Three Fiber (R3F). Every rule below came from a render that was rejected. Deeper notes live in references/:

  • camera-and-editing.md: the beat rig, transitions, long-form vs Short grammar, vertical reframing
  • look-and-ink.md: grade, lights, the post-process ink outline, materials, numbers that worked
  • flicker-and-determinism.md: every flicker cause found, its fix, how to hunt the next one
  • render-and-qa.md: render recipe, motion blur, the GPU lock, QA tools and thresholds
  • shorts.md: Shorts from scratch and Shorts cut from a long form, captions, sound-off readability
  • audio-and-voice.md: voice-over workflow, music, SFX levels, loudness
  • production.md: plan-first workflow, gates, contracts, working with sub-agents and budgets, privacy
  • characters.md: one character family, form shading, 3/4 heads, character ink, acting
  • sets-and-materials.md: painted textures, per-material recipes, set dressing, real scale, sky, sea and weather

scripts/ (Python 3 + numpy + ffmpeg, paths as arguments): flicker.py (one-frame spikes + broken-tab stride test), analyze_pacing.py (beats/min, % still, % fast, hard cuts, mean L*), gpu_lock.py (one render on the GPU at a time).

0. Start every new project from templates/

templates/ is a Remotion 4 starter with the whole look already wired. For a new project:

  1. Copy the contents of templates/ into the new project folder and run npm install (the one install this skill asks for; on the first render Remotion itself also downloads its Chrome Headless Shell, ~130 MB).
  2. Render the demo still first: npx remotion still src/index.ts Demo out/demo.jpg --scale=0.5 --gl=angle --timeout=240000. Look at the DemoHero, DemoWide and Demo-<mood> compositions before building a set, so you know the baseline.
  3. Build on it; don't rebuild what is there:
    • the post chain (ink, grade, bloom, DOF), toon and painted materials, moods (day, dusk, night, cold, lamp, golden), per-character cast lights and the shadow PCF patch;
    • the beat camera rig (aim, poseAt, shot presets in core/shots.ts);
    • IK (solveArm, handWorld, holdInFront) for props held in the hand;
    • all-white captions; a thick-liquid material (core/liquid.tsx);
    • motion blur (DemoMoveSub + ffmpeg tmix, render-and-qa.md); the word-anchored planner plan/plan_src.py;
    • a finished character system, src/inkbeat/char/Character3D.tsx, with one default family (char/family.ts: every proportion, palette and face-style number) and three example characters (char/cast.ts: an adult, a kid, an elder). Design your own cast by changing the overrides; keep every character of a film in one family (characters.md). DemoCast shows the three in one light.

For a target look, give Claude 5-10 frames from videos you like, for private analysis only: measure light, color and pacing, but don't copy their characters, designs, sets or other distinctive elements. Keep the frames outside the project and its git repo, and never ship or publish them.

1. Two grammars: pick the right one

Shorts (vertical), the viral 3D cartoon cadence (measured on a sample of popular 3D cartoon Shorts):

  • There are no cuts, there are camera beats. About 56 beats/min (range 45–66), a new framing every ~0.86 s, hard cuts ≈ 0 per video. The move is the transition: dolly, orbit, crane, whip, push-through, dissolve for time jumps.
  • The camera is never still: ~5% still frames, ~34% fast-motion frames. Between moves the camera keeps drifting at about a third of peak speed. Moves last ~19 frames with a symmetric ease-in-out and start ~6 frames before their word.
  • Pushes in outnumber pulls out about 2:1. Small rolls inside orbits. Shake only on a comic impact.
  • The key object is on screen in the first 2 s. The pace is fastest in the first 40% and ~25% calmer for the payoff; the explanation gets the longest, calmest holds.
  • Sentence starts land on a beat (~68%). Moves follow clauses, not every word.

Long form (16:9), a calm film: the Shorts cadence on a long film reads as cheap and tiring.

  • Hold each framing ~3–6 s (inserts such as a clock, phone or calendar ≥ 2 s), cut with classic coverage (master → medium → close, reverses, keep the 180° line). No orbits, swings, whips, hand-held drift, decorative moves, or inserts that cut away and straight back.
  • A few zooms: a slow eased push across a whole hold on about 1 framing in 3, 0.06–0.14 of the lens-to-subject distance.
  • No burned-in captions in long form (ship an SRT). Held frames are perfectly still (drift 0): see section 5.

2. Plan first, never type a time

  • Get word timings for the final voice-over (TTS alignment or forced alignment). Write a small plan_src.py that anchors every setup and beat to (turn, "phrase", occurrence) and generates plan.json + the TS module your composition imports (the starter plan/plan_src.py does this for shots, captions and tags; add beats and a readable PLAN.md as you grow it). To move a beat, edit the source and re-run.
  • Print the stats every run: beats/min, longest holds and where, hard cuts, sentence starts on a beat, an estimate of fast-motion share. Targets for Shorts: 56 ± 4 beats/min, ≤ 35% fast frames, ≈ 0 hard cuts. Planning ~10% more beats than the target helps: small moves on flat frames don't register on the analyzer.
  • Light follows the acts: give every act its own light and palette (dusk, night, a cold interior, or a golden payoff).
  • Put sound-off tags in the plan (section 7). Work in gates: plan → STOP → look-dev stills → STOP → pilot/Short → STOP.

3. Look: crisp, not milky

Haze, wide bloom, heavy depth of field, grain and lifted blacks read as cheap. Target:

  • a clear key light with definite shadows; heads with a lit and a shadow side (never flat white);
  • a saturated clean palette per act; bloom only on practical light cores; mild DOF; no grain;
  • bright day frames, mean L* ≈ 50–54; slightly lifted, tinted blacks are fine for a Short, true black ink on top;
  • premium pass: stronger key and rim, less ambient, a fresnel rim term in the toon material, soft halation on highlights; outdoors, aerial haze and sun rays in the distance only (haze on the subject washed the opening out). Eye-check the first second after every light change.
  • Unlit colors > 1.0 burn to cream in the grade: keep emissive cores ≤ 1.0–1.2 and saturated.
  • Ink outlines in post, not inverted hulls on everything: depth + normal + object-id edges; ink ≥ 2.5% of the frame, colored per mood, never neutral black. Characters get heavier ink than the set. Settings in look-and-ink.md.
  • Warm against cool in every frame, no monochrome washes, no stacked shade tints; shadows colored, never gray; night is luminous ultramarine, not murky (numbers in look-and-ink.md).
  • Painted textures (texture energy ≥ 0.45, but flat surfaces quiet so they don't read as dirt), ≥ 12 props in every framing, real scale, toy-like props (sets-and-materials.md).
  • Characters: one written-down character family, lit and shadow side, 3/4 heads, always acting, never a real person's likeness (characters.md).

4. Physical plausibility (check every framing)

  • Hands and arms vs blankets, tables, laptops, books, mugs; doors vs heads; bodies vs beds, walls, chairs; walks go through open doorways. If a pose can't avoid it, change the pose, blocking, prop size or framing.
  • Props at real scale (a book ≈ 0.6 × head width), held IN the hand with two-bone IK to a world target, never floating.
  • A floating UI panel is anchored to something (a desk, a wall, a prop); never alone in the air.
  • Named actions are shown and readable: if the voice says someone lifts, opens or climbs something, show the whole action.
  • The camera is never inside geometry (check the whole move, not only the keys), never above a missing ceiling, never beyond a set edge (an interior must never show sky). Nothing sits where a character's head overhangs its shoulders.
  • Moves between distant framings must stay framed the whole way (no empty sky mid-move): pan from one position or route the move past something.

5. Flicker: the causes found, and the rule behind each fix

  1. Per-tab state. Remotion renders with several browser tabs; any three.js state initialized once per tab differs between tabs. Every frame must be a pure function of the frame number: no Math.random, Date, performance.now, no one-time-initialized objects whose state changes later, no caches keyed by Object3D.id, nothing mutated only on some frames. Example: a shadow camera whose projection was computed once → recompute and call updateProjectionMatrix() every render.
  2. GPU overload. Three renders in parallel × concurrency 8 lost GL state: every 8th frame went dark. One render on the GPU at a time (a lock, see render-and-qa.md); concurrency ≤ 8 for finals, ≤ 3 for tests.
  3. Camera drift makes screen-space ink and bloom crawl on fine detail → long-form held frames are perfectly still; Shorts move but use motion blur.
  4. Dotted ink: double-sided wavy planes (curtains) → one flat normal per double-sided mesh in the normal pass; object boundaries → object-id edges; sub-pixel sticks → skip ink below a projected thickness; grazing planes → turn off the one-sided depth-jump test; animated grain or ink wobble → 0.
  5. Z-fighting: coplanar overlapping boxes (wall corners from two boxes sharing a face, ground slabs whose tops meet) → offset ≥ 1 cm, start the second wall behind the first; decals ≥ 3–4 mm proud; live mirrors off when unused. Bars and frames must only touch, never overlap (overlaps inked a flickering black blob).
  6. Shadow crawl: snap the shadow box to whole shadow texels and quantize its size; patch three's screen-locked PCF dither (it crawls under any camera move).
  7. Memory: a per-object material cache grew every shot → the browser ran out of memory. Bound every cache.
  8. Strobing: fast periodic motion (anything cycling at ~4 Hz) and very fast lens moves read as flicker even with motion blur: cycles ≤ ~2 Hz, big moves ≥ ~0.85 s. Hunt the next one with scripts/flicker.py, then crop the region over 3 frames (flicker-and-determinism.md).

6. Render and QA (summary; recipe in render-and-qa.md)

  • WebGL in headless Chrome: --gl=angle, a long --timeout (240 s), one render process, one bundle per project (bundles copy public/: clean up old ones you created, watch the disk).
  • Motion blur on every Short: render 4 sub-frames per frame across a 180° shutter at 4× fps, average with ffmpeg tmix, keep every 4th frame. Captions are keyed to the output frame, so they stay sharp.
  • QA on every render: flicker.py (spikes from whips or lightning are fine; trace the rest), analyze_pacing.py (beats/min, % still, % fast, cuts, L*), a 2 fps contact sheet eye-checked for interpenetration and empty frames, loudness, and a muted read of the sheet: does the story read without sound?

7. Shorts specifics (more in shorts.md)

  • Captions: 1–2 lowercase words, swapped exactly on word onsets (2 frames early), no pop or scale, centered at 70% height, a heavy geometric sans, ≥ 90 px at 1080 wide, a heavy dark outline, every word white (an accent color is optional). Spoken punctuation kept.
  • Sound-off readability: viewers scroll muted. Put the story in the picture with 1–3 word tags painted on the key objects (a year, a price, a label), bold, large, high contrast, unlit so the light can't wash them out.
  • Vertical framing: subject in the upper part (faces at 30–45% height) so captions at 70% never cover faces.
  • Zoom levels: the push inside a hold ~3%/s, on a set change ~12% (7% or 30% read as too busy).
  • The picture ends ~0.7 s after the last word.

8. Audio (more in audio-and-voice.md)

  • Music bed ~14 dB under the speech (long-form edits ~20+ dB), a −3 dB duck under the voice, faded at the end.
  • Story SFX on the exact word, 6–8 per 20 s, each leveled against the voice (peak ~8 dB under the voice peak, the loudest 100 ms ≥ 6 dB under the speech RMS); one global SFX gain lets transients and sustained tones jump out.
  • Two-pass loudnorm to −14 LUFS, −1 dBTP. Never overlap-crossfade cut VO segments (it shortens audio, captions drift).
  • Check licenses of every music track you ship; a library track can still draw a content claim.

9. Production habits that saved the most time

  • A contract file + plan JSON per sequence + one owner per area (core, sets, cast, sequences, UI, audio).
  • A pilot (~80 s or the first 15 s of a Short) under the new rules, shown before rolling out.
  • Shared sets change only additively: a new variant flag that defaults to the old behavior.
  • A typecheck gate before every bundle (one syntax error in an auto-loaded module breaks everyone's bundle).
  • Sub-agents burn budget fast (many in parallel can use up a usage window in well under an hour): default to none, at most 2–3; at most 3 images per step.
  • If asked to work autonomously, don't stop for gates. Confirm a defect still exists before fixing it.
  • Nothing from the maker's computer or identity in anything published (paths, user names, emails, dates); no real brands or logos (production.md).

관련 스킬