Declarative motion graphics that AI can write, humans can tweak — and the human's edits survive an AI regeneration.
At "prompt → mp4", reframe is on par with Hyperframes or a Remotion skill — our own benchmark says exactly that, parity not superiority. The difference starts on the second turn: what you get back is not freeform code but an addressable document, so you can keep tweaking, regenerating, and scaling it without your changes being silently lost.
Top: a scene with human overlay edits applied (brand color, retimed reveal, watermark). Bottom: an AI fully regenerated the base scene — different layout, different timing — and the same overlay reapplied by stable id. 4 edits survived; the one renamed node was reported as an orphan, never silently dropped.
The demo above is a reframe scene
(examples/scenes/reframe-demo.ts) —
render it yourself: pnpm reframe render examples/scenes/reframe-demo.ts.
scene.ts ──(written by you, or by an AI given `pnpm reframe guide`)──▶ IR (plain JSON data)
│ │
▼ ▼
preview: scrub + knobs ──▶ edits recorded as an overlay JSON (non-destructive)
│ │
▼ ▼
render: deterministic mp4 (same input → byte-identical frames) ◀── overlay reapplies
even after an AI
regenerates the base
Everything is a pure function of time: evaluate(scene, t) — no wall clocks,
no randomness without a seed, scrubbing and distributed rendering for free. And
it's enforced: reframe lint compiles the scene twice and flags any IR that
differs (a Math.random() or Date baked into a prop), so a scene that would
silently render differently each time fails the gate before you ship it.
The kind of scene that is hand-rolled timeline math in GSAP or per-element
interpolate() plumbing in React is ~100 lines here, because the host
language generates the nodes, states, and phase-shifted behaviors — and the
output is still data: every dot, glyph, and moon keeps a stable id you can
tweak in the preview.
![]() bloom.ts — 300 dots on a golden-angle spiral: radial bloom, traveling breath wave, chromatic ripple, vortex collapse |
![]() wavefield.ts — physical interference: 1,152 phase-shifted oscillators on a 32×18 grid, second ripple source joins mid-scene |
![]() orbit.ts — nested transform composition: moons orbit planets orbit a sun, the whole system tilts — three nested groups, three linear tweens |
![]() typewave.ts — character-level kinetic type: cascade, standing wave, shatter with spin, and a second phrase assembling from the debris |
![]() glyph-reveal.ts — the archival stop-motion format: AI-generated plates as image nodes, ~7fps hard cuts, push-in, camera shake, a tick per cut. Swap any plate from an overlay or batch row: nodes.frame-3.src |
Motion is hard to put into words and tedious to hand-key, so reframe ships a
small set of named motion presets: draw-bloom, punch-in, rise-settle,
slide-bank, reveal-orbit, spin-forge. Each is a seeded generator, not a
canned template. The same name yields a family of distinct takes, and a seed
varies it within that family (an 8-seed spread stays measurably different yet
recognizably the same motion, gated in the tests).
motionPreset("spin-forge", { target: logo, energy: 0.8, seed: 3 })energy (clean to springy) and speed are universal knobs. The preset emits a
beat you can retime, and a waypoint you drag on its path becomes an overlay edit
that survives a knob-driven regeneration: knobs regenerate the base, hand craft
persists. Two primitives back it:
pathnode: a true vector SVG shape with aprogressdraw-on, so the outline draws itself, stays crisp at any zoom, and recolors by animating fill.motionPath: drives a node's x/y along a Catmull-Rom curve (with tangentautoRotate), the curved motion straight tweens cannot do.
Turn any SVG into a share-worthy animated sting in one command, no clone needed (a local file, or any of simple-icons' brands):
npx reframe-video logo react --motion spin-forgeSix real logos (React, Figma, Vercel, GitHub, Notion, Stripe), each a different
preset, one command each. Pulled from simple-icons by slug, or drop in your own .svg.
Promo and product shots almost always need the app inside something — a phone, a
browser, a laptop. The tedious part was never the frame, it was getting the
screen clip right so content sits inside it. devicePreset bakes that in:
ten parametric vector frames, each with a clipped screen "content slot", so a
mockup is a single call.
devicePreset("phone", { id: "hero", content: [ /* ...your UI nodes */ ] })Ten frames — phone, tablet, laptop, browser, watch, monitor, tv, foldable,
terminal, car — all pure primitives (no assets), deterministic, and additive
to the golden contract. deviceScreen, deviceScreenCenter, and deviceBounds
hand back the screen bounds and the frame footprint, so content scrolls inside
the clip and many devices tile onto a grid. The motion is yours: drive the
device group with tweens or a motionPreset, and the content scrolls because it
is clipped. See examples/scenes/device-presets.ts (three devices, a scrolling
feed) and device-teardown.ts (all ten, each with its own signature move).
Because their output is arbitrary HTML/React — great to generate once, impossible to safely operate on afterwards. reframe's output is data with stable addresses, and everything below falls out of that one difference:
| the second turn | HTML / React output | reframe |
|---|---|---|
| "tweak just the color and timing" | edit code by hand, or re-prompt and hope nothing else changes (no visual editor is possible over arbitrary code) | turn knobs in the preview — no code |
| "now redesign it" after my tweaks | your hand edits live inside the code; regeneration overwrites them or you merge diffs — silent loss is the default | edits live in an overlay; they reapply onto the regenerated scene (measured 100% across 23 regenerations/turns), breaks are reported loudly |
| "make 50 personalized versions" | only what the author pre-parameterized (props) | any address, post-hoc: nodes.name.content |
| "is it wrong before I render?" | semantic failures are invisible until pixels (wrong text, off-frame) | structure validates pre-render with actionable errors; motion is computable from the IR |
If your video is fire-and-forget, use the simpler tool. If it's an asset that will be tweaked, regenerated, and multiplied — that loop is what reframe is for.
No clone needed — reframe-video is on npm:
brew install ffmpeg # system dep (or apt install ffmpeg)
npx playwright install chromium # one-time browser download
npx reframe-video new hello # scaffold hello.ts in any directory
npx reframe-video render hello.ts # → out/hello.mp4Using Claude Code? Install the skill and just describe the video you want:
/plugin marketplace add kiyeonjeon21/reframe
/plugin install reframe@kiyeonjeon21
To hack on reframe itself, clone-based setup:
brew install ffmpeg # 0. system dep (or apt install ffmpeg)
pnpm install # 1.
pnpm exec playwright install chromium # 2. one-time browser download
pnpm reframe render examples/scenes/lower-third.ts # 3. → out/lower-third.mp4Then open the editor and render with your edits:
pnpm reframe preview # scrub, play, and edit any scene with knobs
# → edits accumulate in an overlay; click "download", save to examples/overlays/
pnpm reframe render examples/scenes/logo-reveal.ts \
--overlay examples/overlays/brand-edits.jsonSaving edits: the preview never modifies your scene file. Edits live in an overlay document — download it from the panel, then pass it to render with
--overlay. That same file keeps working after the scene is redesigned.
pnpm reframe demo # renders out/demo-{1,2,3}-*.mp4Renders the base scene, the base + a human overlay, and then an
AI-regenerated base + the same overlay. Watch the console: surviving edits
apply, the deliberately renamed node orphans loudly with a diagnosis. This is
the project's core claim, reproducible in one command. (Measured with real
agents: 100% id/state/label retention across 8 regenerations —
benchmark/regen/REGEN-ANALYSIS.md.)
Because choreography is data, audio cues anchor to timeline labels, not to seconds on a waveform:
audio: {
bgm: { synth: "ambient-pad", gain: 0.3, duck: { depth: 0.5 } },
cues: [{ at: "enter", sfx: "whoosh" }, { at: "shatter", offset: 0.18, sfx: "thud" }],
}Retime a step with an overlay — or let an AI regenerate the scene — and the
sound design moves with it (verified: a +1.8s hold patch shifted the anchored
cues by exactly +1.8s). SFX are procedurally synthesized (deterministic, zero
assets) with CC0 samples in assets/sfx/ for organic sounds like real
mechanical keypresses; the bed auto-ducks under cues. --no-audio to skip.
A scene is a template; every data row becomes an overlay. Row keys are overlay addresses — no new schema:
pnpm reframe batch examples/scenes/lower-third.ts examples/data/team.json
# → out/batch/alice.mp4, ben.mp4, ... + batch-report.jsonRows render in parallel; a row with a bad address renders with a loud orphan warning instead of killing the batch. CSV works too (headers = addresses). This is N-personalized deterministic videos from one template — the workflow real-time runtimes like Rive structurally don't cover.
import { scene, text, seq, to, wait } from "@reframe/core";
export default scene({
id: "hello",
size: { width: 1920, height: 1080 },
fps: 30,
background: "#101014",
nodes: [
text({ id: "title", x: 960, y: 540, anchor: "center",
content: "Hello", fontFamily: "Inter", fontSize: 120, fontWeight: 800, fill: "#FFF" }),
],
// base props are the finished design; states are sparse overrides
states: {
hidden: { title: { opacity: 0, y: 580 } },
shown: { title: { opacity: 1, y: 540 } },
},
initial: "hidden",
// labels are stable addresses for overlay timing edits
timeline: seq(
to("shown", { duration: 0.6, ease: "easeOutCubic", label: "enter" }),
wait(2, "hold"),
),
});Scaffold one with pnpm reframe new my-scene. Full syntax (node types,
states, timeline operators, behaviors): pnpm reframe guide — the guide an
LLM reads to write valid scenes on the first try (33/33 first-attempt renders
in our benchmark).
A scene is a single self-contained file, not an app: it can live in any
directory — no package.json or node_modules next to it. render bundles it
on the fly (resolving @reframe/core and any relative imports beside it), and
preview lists scenes from the directory you launched it in alongside the
repo's examples/scenes/. Keep overlays and batch data files right next to
your scene.
| command | what it does |
|---|---|
pnpm reframe render <scene.ts|.json|.html> [--overlay f]... [-o out] |
deterministic mp4 (mode inferred from extension; output defaults to out/) |
pnpm reframe batch <scene.ts> <data.json|csv> [-o dir] [--overlay f]... |
one mp4 per data row (rows = overlays), parallel, with a per-row report |
pnpm reframe compile <scene.ts|.json> [-o out.json] [--json] |
bundle + validate a scene to SceneIR JSON, no render (fast; no ffmpeg/chromium); --json returns {ok, kind, issues} |
pnpm reframe frame <scene.ts|.json> [--t <sec>] [-o out.png] |
render one frame at time t to a PNG (chromium only, no mux) — for a render-and-look loop |
pnpm reframe manifest <scene.ts|.json> [--json] |
dump the addressable surface: every node, state, timeline label, and beat with the overlay address that reaches it |
pnpm reframe lint <scene.ts|.json> [--json] [--strict] |
the studio-readiness gate: flag un-addressable motion + verify the scene is a pure function of time; --strict exits non-zero |
pnpm reframe verify-overlay <base> <overlay>... [--json] |
compose an overlay onto a base and report applied-vs-orphaned, no render — the regen-survival check (non-zero exit on orphans) |
pnpm reframe labels <scene.ts|.json> |
print the compiled event clock (every timeline label → exact seconds) — the timing source for audio cues |
pnpm reframe assemble <media...> [-o name] |
probe images/videos (ffprobe) and scaffold an editable montage scene .ts wired with photoMontage |
pnpm reframe player <scene.ts|.json> [-o out.html] |
bundle a scene into one self-contained HTML that plays the motion live in any browser |
pnpm reframe logo <logo.svg|brand-slug> [--motion <preset>] |
animate a logo (or a simple-icons brand) into a sting |
pnpm reframe diff <ref-image> [scene.ts] [--t <sec>] [--mode side|blend|diff|grid] |
compare a render against a reference image |
pnpm reframe preview |
scrub/play/edit UI; edits export as overlay JSON |
pnpm reframe new <name> |
scaffold a documented starter scene |
pnpm reframe motion <mp4|framesDir> |
calibrated motion profile (speeds, easing, discontinuities) |
pnpm reframe trace <ref.mp4> [--apply scene.ts] |
extract a video's motion structure (a MotionSketch); --apply emits a timeline that re-tells it on your own nodes |
pnpm reframe guide [--directing|--regen|--html] |
print a guide: eDSL syntax (default), the high-end directing workflow, the regeneration contract, or HTML/GSAP scenes |
pnpm reframe skill [--path] |
print the authoring skill for an agent; --path prints the plugin dir to load |
pnpm reframe demo |
the edit-survival demo above |
Overlays address the scene by node id, state name, and timeline label (or
beat name) — never by position or index. When an AI regenerates a scene it follows one
contract (docs/guides/regen-contract.md, or pnpm reframe guide --regen): keep
those names stable for every concept that survives the redesign. When the
contract is broken anyway, composeScene skips the affected edits and reports
them with a diagnosis naming the likely rename. The failure hierarchy:
- Contract followed (the measured common case) → edits survive.
- Contract broken → loud orphan report.
- Never: silent edit loss, or a render failure caused by base drift.
An overlay isn't limited to patching props and timing — it can change the
structure of the cut, keyed by the same stable addresses, and those edits
survive regeneration too: reorder a beat (timeline.<beat>.order),
remove one (removeTimeline: ["shot-3"]), or insert a whole new unit
(insertNodes + insertTimeline { into: "montage", after: "shot-2" }). A
montage is built so each shot is the self-contained beat shot-${i}, so an
overlay can drop, reorder, or splice in a card without touching the base:
// reorder two cards, drop one, retitle another — all survive an AI regen
{ "removeTimeline": ["shot-3"],
"timeline": { "shot-1": { "order": 2 }, "shot-2": { "order": 1 } },
"nodes": { "shot-0-title": { "content": "REORDERED" } } }Reproduce it with the pure-vector demo (no assets) — the same overlay reorders, removes, and inserts a card, and every edit is reported applied with zero orphans:
pnpm reframe verify-overlay examples/scenes/vector-montage.ts \
examples/overlays/vector-montage-restructure.json # 4 applied, 0 orphaned
pnpm reframe render examples/scenes/vector-montage.ts \
--overlay examples/overlays/vector-montage-insert.json # a new card spliced into the cutThe CLI is a thin shell over an importable, server-side API — load, validate,
and check a scene without spawning a process. From reframe-video/compile:
import { loadScene, loadSceneFromCode, checkDeterminism } from "reframe-video/compile";
const compiled = await loadScene("scene.ts"); // bundle + validate → CompiledScene
// On failure a SceneLoadError carries .kind ("bundle" | "eval" | "validation")
// and .issues: { code, path, message }[] — structured, not prose to parse.
const { deterministic, findings } = await checkDeterminism("scene.ts");reframe-video/renderer exposes renderFrame / drawDisplayList (DisplayList →
Canvas 2D) for live browser playback, and reframe compile --json returns the
same structured shape on the CLI: { ok: false, kind, issues: [{ code, path, message }] }
— the feedback loop an agent or a UI reads to point at the exact broken node.
📖 docs.reframe-video.com — the full documentation
site. The docs/ folder is its Mintlify source (docs/docs.json):
| page | what |
|---|---|
| Introduction · Quickstart · The loop | the pitch, install, and the AI-write / human-edit / deterministic-render model |
| Gallery | a curated visual reel of scenes |
| Examples | all 67 example scenes, by category |
| Guides | the eDSL, directing, HTML/GSAP, and regeneration-contract guides (also pnpm reframe guide) |
Curated renders live in docs/assets/gallery/ and accumulate via pnpm gallery (the committed home; out/ stays scratch).
| path | what |
|---|---|
packages/core |
the eDSL, IR, determinism kernel, overlay composition — zero deps |
packages/renderer-canvas |
DisplayList → Canvas 2D (browser + capture shared) |
packages/render-cli |
Playwright capture + ffmpeg encode; also renders arbitrary HTML/GSAP deterministically via a virtual clock |
packages/preview |
the Vite editor |
examples/ |
67 example scenes (see examples/README.md), overlays, compositions, the edit-survival demo |
labs/ |
experiments and product probes (live-data → baked scene → render), kept out of examples/ so it stays purely demonstrative |
docs/ |
the Mintlify-ready docs site + the authoring guides (also pnpm reframe guide) |
benchmark/ |
measurement artifacts, not product code: LLM generation benchmark (RESULTS/ANALYSIS.md), regeneration-contract experiment (regen/), calibrated motion profiler (harness/motion/, MOTION.md) |
- Node ≥ 20, pnpm ≥ 9, ffmpeg on PATH.
Executable doesn't exist at …/ms-playwright/…→ runpnpm exec playwright install chromium(the workspace blocks postinstall scripts, so the browser is not fetched automatically).spawn ffmpeg ENOENT→ install ffmpeg (step 0).- Fonts: only Inter 400/700/800 are bundled; other families silently fall back.
- A scene importing npm packages beyond
@reframe/coreonly bundles if those packages are resolvable from the scene's directory — scenes are meant to be dependency-free documents.
Early alpha (reframe-video on npm, Claude Code skill in this repo). The
research phase is closed: every design hypothesis is measured, not assumed —
LLM generation parity with HTML+GSAP, deterministic byte-identical rendering,
edit survival across AI regeneration, and a five-turn natural-language
iteration loop with zero silent edit loss. Receipts: benchmark/ANALYSIS.md,
benchmark/MOTION.md, benchmark/regen/REGEN-ANALYSIS.md,
benchmark/nl-loop/NL-LOOP.md. What "alpha" means honestly: it has not met
strangers yet — surface area is intentionally small (8 node types, one font,
Canvas 2D) and the IR/overlay schema has no compatibility promise before 1.0.








