remotion-assemble

How to use Remotion as the assembly layer - read the project's meta.json and stitch the rendered HyperFrames scenes + footage + voice + sound effects into the finished video. Read this when writing/ed

SKILL.md

GitHub

Remotion Assemble - building the timeline from meta.json

In this system Remotion does exactly one thing: assemble the pieces that already exist (scene MP4s rendered by HyperFrames, source footage, voice, sound effects) into the final video. Do not build motion graphics in Remotion - that is HyperFrames' job.

Structure of engines/remotion

engines/remotion/
├── package.json          ← remotion, @remotion/cli, @remotion/renderer
├── remotion.config.ts
└── src/
    ├── Root.tsx               ← registers 3 compositions: "Assemble" + "Poster" + "Thumbnail"
    ├── Assemble.tsx           ← master composition: read manifest → build timeline
    ├── Poster.tsx             ← static poster composition (image projects)
    ├── Thumbnail.tsx          ← static video-thumbnail composition (frame card + title)
    ├── manifest.ts            ← load + validate video props (zod)
    ├── duration.ts            ← scene length incl. `speed` (shared with the web editor; server twin in timeline.ts)
    ├── color.ts               ← timeline color adjust → CSS filter (`colorFilterCss`, `ColorDefs`, preset approximations)
    ├── layerGeometry.ts       ← `layerBox()` - layer position/size/animation/keyframes (shared with the editor's drag frame)
    ├── posterManifest.ts      ← load + validate poster props (zod)
    ├── thumbnailManifest.ts   ← load + validate thumbnail props (zod)
    ├── brandFonts.ts          ← load brand fonts via staticFile (offline)
    ├── media.tsx              ← media URL resolver (staticFile by default; the dashboard preview swaps it)
    ├── index.ts               ← entry point registering Root
    └── components/
        ├── SceneClip.tsx      ← <OffthreadVideo>/<Img> for one scene/footage clip
        ├── Transition.tsx     ← scene transitions inside transitionOverlap (fade, slide, wipe, zoom, blur, none)
        ├── LayerTrack.tsx     ← overlay layers (image / video PiP / text / shape)
        ├── ColorBox.tsx       ← wraps a scene/layer with its color adjust + vignette
        ├── SfxTrack.tsx       ← places <Audio> at sfx[].atFrame
        ├── MusicTrack.tsx     ← background music: loop + fade + deterministic ducking
        ├── CaptionTrack.tsx   ← karaoke captions from word timestamps
        ├── HighlightTrack.tsx ← main/related key bands (key-layout)
        └── vietnameseFont.ts  ← @font-face Inter vietnamese subset

Manifest fields beyond scenes/captions/overlays (see manifest.ts):

  • watermark - the Style Design corner logo, stamped by the SERVER (jobs/assemble.ts), never by the AI; null = the style has no logo.
  • audio.music - background music { file, volume, duckVolume, speech: [[startSec,endSec],...] } with deterministic ducking over speech ranges (see the background-music skill); null = no music.
  • scenes[].zoom - camera move on footage { origin?, keys: [{frame, scale, ease?}] }; frame counts from the START OF THE SCENE, applied to the wrapper div (never the video element).
  • audio.sfx[].mediaStart - seconds skipped at the head of the sfx file (lead silence trim, see noti-tiktok-vn "SFX loudness and lead silence").

Editor v2 fields (all optional - absent = the old behaviour)

Full contract: docs/EDITOR-V2-PLAN.md section 1. The server validator (apps/server/src/timeline.ts) rejects anything out of range with 400 INVALID_TIMELINE, and the zod schema has the same ranges, so an out-of-range value written by hand kills the render. Stay inside these ranges.

Scene (scenes[]):

Field Range Meaning
fit "cover" (default) | "contain" how the source fills the frame
transform { x, y: -1..1, scale: 0.1..5, rotate: -180..180, flipH, flipV } x/y = OFFSET as a fraction of the frame side (x = 1 moves a full width); scale multiplies the existing zoom
crop { top, right, bottom, left } each 0..0.45 crops the SOURCE, then fits the rest (object-view-box, no black bars)
color ColorAdjust (below) color on the timeline - does not touch the file
volume 0..2 (footage) muted: true still wins
fadeInSec / fadeOutSec 0..10 (footage) audio fade of the clip's own sound
speed 0.25..4 (footage) playback rate. Length = (round(to·fps) − round(from·fps)) / speed when durationInFrames is absent - never compute it yourself with a different formula; the helper is duration.ts (server twin sceneDurationFrames). durationInFrames WINS over speed: if the scene has one, rewrite it in the same edit (and pin to) or the clip ends early (black tail) / gets cut
transition { type } transition INTO THE NEXT scene, played inside this scene's transitionOverlap. Types: fade (default, the old crossfade) slide-left slide-right slide-up slide-down (push) wipe-left wipe-right zoom-in zoom-out blur none. No overlap = hard cut whatever the type

ColorAdjust (scene color and layer color): preset = one of the 14 ids in apps/server/src/color.ts (approximated with CSS), brightness contrast saturation temperature tint -1..1 (0 = unchanged, saturation -1 = black and white, temperature + = warm), hue -180..180°, vignette 0..1, blur 0..20 px. This is a look on the timeline, not the ffmpeg asset grade of the color-grading skill - for HDR/log footage still delog the file first.

Layers (top-level layers[]) - drawn above the scenes, below highlight/captions/watermark; later entries are on top:

"layers": [
  { "id": "title", "kind": "text", "from": 30, "durationInFrames": 90,
    "text": "Dòng 1\nDòng 2", "width": 0.8, "transform": { "y": -0.6 },
    "style": { "fontSizePx": 84, "color": "#ffffff", "backgroundColor": "rgba(0,0,0,0.6)", "paddingPx": 24, "radiusPx": 20 },
    "animation": { "in": "pop", "out": "fade", "durationSec": 0.4 } },
  { "id": "pip", "kind": "video", "src": "assets/broll.mp4", "mediaStart": 2,
    "from": 120, "durationInFrames": 150, "width": 0.4,
    "transform": { "x": 0.5, "y": 0.55 }, "radiusPx": 24, "muted": true },
  { "id": "logo-openai", "kind": "image", "src": "assets/openai.svg", "aspect": 1, "from": 0, "durationInFrames": 60,
    "keyframes": [ { "frame": 0, "scale": 0.8, "ease": "out" }, { "frame": 20, "scale": 1 } ] }
]
  • kind: image | video (both need src, a project-relative path like every other media path - staged by jobs/assemble.ts) | text | shape (a colored box: style.backgroundColor, aspect w/h default 16/9).
  • from / durationInFrames are ABSOLUTE frames, like cues. keyframes[].frame is RELATIVE to from.
  • Position: transform.x/y is the CENTRE of the box, -1..1 from the frame centre (0,0 = middle, 1 = edge) - different from a scene's transform. width is a fraction of the frame width (0.05..2, default 0.5); height follows the file's aspect (image/video), aspect (shape) or the text.
  • Text: style.fontFamily is one of the SubtitleTrack ids (vietnamese default, sans, serif, mono) - Vietnamese diacritics are safe; fontSizePx is in composition pixels.
  • animation.in/out: none fade slide-up slide-down slide-left slide-right pop zoom; durationSec default 0.4 (capped at half the layer on short layers).
  • mediaAspect (layer) and mediaDurationSec (sfx) are filled by the server/preview at build time (ffprobe) - never write them into meta.json. ffprobe often cannot measure an SVG (depends on the ffmpeg build), so for an SVG image layer set aspect (w/h) yourself, otherwise the box falls back to 16/9.
  • Geometry is computed by layerBox() in layerGeometry.ts; the editor's drag frame uses the same function, so never re-implement it in a component.
  • Logo rule still applies: a brand logo layer must come from assets/brand-logos/ copied into the project - and never add the Style Design corner logo as a layer (the server stamps it).

Audio fades: audio.sfx[].fadeInSec/fadeOutSec (0..10, fade-out counts back from the END of the file), audio.music.fadeInSec/fadeOutSec (override the 0.5 s / 1 s defaults; 0 = no fade).

Core principles

  1. A single data-driven Assemble composition. Everything comes from the project's meta.json, passed through defaultProps/inputProps. Adding a new video means no Remotion code changes, only a new manifest.
npx remotion render Assemble \
  --props="<abs>/video-projects/<name>/props.resolved.json" \
  --output="../../outputs/<name>-v1.mp4" \
  --concurrency 8 --gl angle

⚠️ Asset paths in the props must be staging/... - the backend stages assets into engines/remotion/public/staging/ via hardlinks and then writes props.resolved.json. NEVER render straight from meta.json (its paths are relative to the project folder, and Remotion cannot resolve those through staticFile).

--gl angle is MANDATORY on a machine with a GPU - without it Remotion renders with the software renderer, CPU at 100% and GPU at 5%. For drafts add --crf 28 --x264-preset veryfast. (The backend queue adds these flags automatically.)

  1. Dimensions/fps come from the manifest - use calculateMetadata to set width/height/fps/durationInFrames dynamically from the props, never hardcode them in Root.tsx.

  2. Embedded video uses <OffthreadVideo>, not <Video> - it is more stable server-side and lands on the right frame. Whatever fps a HyperFrames scene was rendered at, the manifest must declare that exact fps; an fps mismatch between scene and composition is the number one source of stuttering.

  3. Timeline = accumulating durationInFrames:

let from = 0;
scenes.map((s) => {
  const seq = (
    <Sequence key={s.id} from={from} durationInFrames={s.durationInFrames}>
      <SceneClip scene={s} />
    </Sequence>
  );
  from += s.durationInFrames - (s.transitionOverlap ?? 0);
  return seq;
});

When a transition has an overlap, subtract that overlap while accumulating - forgetting to subtract leaves a black gap between scenes.

  1. Audio:

    • Voice: one <Audio src={voice}> running from frame 0 - the voice is the sync backbone, scenes follow the voice and never the other way round.
    • Sound effects: each sfx[] entry becomes one <Sequence from={atFrame}><Audio volume={0.3}/></Sequence>. Default sfx volume is 0.3 (~10dB below the voice); override it in the manifest with the volume field if needed.
    • Do not normalize/mix manually with FFmpeg after rendering - mix inside Remotion so the draft sounds like the final.
  2. Caption and overlay tracks live in meta.json, and their timestamps are in FRAMES.

    CaptionTrack and HighlightTrack read props.captions / props.overlays only. They never open a transcript file. If you leave those arrays empty there are no captions, no matter how good the transcript is.

// meta.json
"captions": [
  { "from": 90, "durationInFrames": 75,
    "words": [ { "text": "Trí", "start": 90, "end": 96 },
               { "text": "tuệ", "start": 96, "end": 104, "hi": true } ] }
],
"overlays": [
  { "from": 0, "durationInFrames": 150, "tier": "main", "accent": "hot",
    "kicker": "AI", "parts": [ { "t": "Dựng video" }, { "t": "bằng AI", "hi": true } ] }
]

⚠️ The unit changes here. A transcript stores start/end in seconds; captions[].words[] stores them as absolute composition frames (second * fps, counted from the start of the whole video, not from the start of the cue). Passing seconds through produces captions that all flash in the first half-second — and nothing errors, so it is easy to ship by accident.

hi: true marks a highlighted word. tier: "main" renders a top band, "sub" a bottom pill that lifts itself when captions are present.

Any cue (caption, subtitle, overlay) may carry "transform": { "x", "y", "scale" } - an offset from the track's default anchor, x/y as a fraction of the frame width/height (-1..1), scale 0.3..3. The user sets it by dragging the cue on the editor's preview. Leave it alone when you rewrite a cue's text or timing; only change it when asked to move or resize the text.

  1. Every scene needs durationInFrames unless it is a srcVideo scene with both from and to. That is the only case Remotion infers a length for; an image or HyperFrames scene without an explicit durationInFrames throws at render time. This bites hardest on videos with no source footage (text-to-video, image-only explainers), where nothing has a natural duration to infer from.

  2. meta.json is shared with a human editor. The dashboard's video editor previews the Assemble composition live in the browser (@remotion/player) and saves the timeline keys (scenes, audio, captions, subtitles, subtitleStyle, overlays, layers) straight back into meta.json. When you assemble or adjust the timeline: re-read meta.json right before each change, change it in place with Edit, never Write the whole file from memory, and keep every field you do not recognise (the schema uses z.looseObject precisely so extra fields survive). Paths stay relative to the project folder with no .. - the editor's server-side validator rejects anything else, and so does staging.

  3. Asset paths: Remotion code only loads through staticFile() - the backend stages assets into engines/remotion/public/staging/<project>/ (hardlink) and writes the staging/... path into props.resolved.json; Remotion never reads an absolute path. This runs on Windows - always use path.join on the backend, never string concatenation.

    Inside the Assemble tree, get that staticFile through useMediaSrc() (src/media.tsx), never by calling staticFile() directly. The dashboard editor runs the same Assemble in @remotion/player, where there is no Remotion public/: it wraps the tree in MediaResolverProvider and maps project-relative paths to /media/video-projects/<id>/... and fonts/... to /media/remotion-fonts/.... A direct staticFile() (or one evaluated at module scope, like the old vietnameseFontFaceCss constant) renders fine through the CLI but 404s in the preview. The default resolver IS staticFile, so CLI renders produce the same frames as before. Poster/Thumbnail/brandFonts only render through the CLI and keep staticFile.

Draft vs Final

Draft Final
Command --crf 28 --x264-preset veryfast (+ --concurrency 8 --gl angle) defaults (crf 18)
Scene input renders/*.draft.mp4 renders/*.mp4 (quality standard)
Purpose review pacing, sync, transitions publishing

The backend picks the scene input set based on the job type - the Remotion code does not distinguish draft from final, it only takes the paths from the manifest.

Known issues & how to avoid them

  • Stutter/freeze at a scene boundary: the scene fps differs from the composition fps, or durationInFrames in the manifest does not match the real length of the MP4. Check with ffprobe -show_streams <file> before assembling.
  • Audio drifting out of sync toward the end: a VBR voice mp3 -> convert it to CBR/WAV before putting it in the manifest (ffmpeg -i voice.mp3 -ar 48000 voice.wav).
  • Color mismatch between HyperFrames scenes and footage: both engines render through Chromium so they usually match; if they do not, check whether the footage has an unusual color space tag (bt709 is the standard) - transcode the footage to bt709 first.
  • Render hanging on Windows: usually caused by Vietnamese characters/spaces in an asset file path - name asset files in ASCII kebab-case right from the import step.
  • One black frame at every footage cut: a srcVideo scene sized round((to-from)*fps) while OffthreadVideo plays round(to*fps) - round(from*fps) frames - the two differ by one (from=1.02, to=3.98 @30fps: 89 vs 88). Always round EACH END, never the difference; resolveSceneDurationInFrames and frameOf in jobs/assemble.ts must stay in lockstep with SceneClip.
  • Sped-up footage (speed > 1) going black for the second half of the scene: since Remotion 4.0.537 (the pinned version) the Sequence inside OffthreadVideo already divides (endAt - startFrom) by playbackRate itself, so endAt must be the plain SOURCE frame round(to * fps). Converting it again (startFrom + ceil((endAt - startFrom) / speed), which 4.0.500 needed for speed < 1) divides twice: speed 2 on a 10 s clip gives a 150-frame scene whose video stops at frame 75. Check node_modules/remotion/dist/cjs/video/OffthreadVideo.js again whenever you bump Remotion.
  • Whole render dies on one tiny cue (inputRange must be strictly monotonically increasing): a fade written as [0, 4, max(5, last-4), last] is not increasing when a cue is 6 frames or less (a translated cue clamped near the end of the video is enough). Use fadeInOut() from components/fade.ts for any new track instead of hand-writing the range.
  • Rendering on a machine with no network: Remotion tries to download Chrome from remotion.media. Pass --browser-executable=<local chrome/headless_shell> instead.

Remotion license (keep in mind when scaling up)

Remotion is free for individuals and companies of <= 3 people; beyond that you need a Company License. This note is for planning growth - it does not affect the current stage.