Install
That writes one file:compositions/components/motion-blur.html.
Source
motion-blur.html
motion-blur.html
<!--
Motion Blur — After Effects-style shutter-based motion blur.
Usage: paste this snippet into your composition, then put
data-hf-motion-blur on any element your GSAP timeline animates.
Nothing else to call, and no order to get wrong.
How it works — the same model After Effects uses for the layer Motion Blur
switch (temporal supersampling), synthesised inside the page because the
renderer captures one instant per frame with no shutter:
1. After every timeline update (each rendered frame), the snippet seeks
the timeline to `samplesPerFrame` sub-frame times spread evenly across
the shutter window, reads each target's RESOLVED `transform` at every
sample, then restores the frame time. Seeks run with events suppressed,
so no user callbacks fire and the DOM ends up exactly where the frame
left it. Reading the resolved matrix rather than a list of named GSAP
properties is what makes every driver work: translation, scale,
rotation, 3D rotation and skew arrive in the same one value, and a
driver nobody thought of needs no change here.
2. One copy of the element per sample is stacked in an isolated group,
each carrying its sample's transform, each at 1/N opacity with
`mix-blend-mode: plus-lighter`. plus-lighter adds premultiplied colour,
so N+1 copies at 1/N sum to the average — the shutter integral. The
isolation is load-bearing: without a transparent backdrop of its own
the first copy would add onto the page and blow a light background out
to white. Result: a smear along the actual trajectory whose extent is
speed × shutter time, with no decaying one-sided trail. The tails —
where fewer copies overlap — step down evenly at 1/N per copy; wherever
the element travels less than its own size the copies pile up past 1
and clamp, so the middle of the smear is solid rather than a plateau.
3. The element itself paints over that group, sharp and at full opacity.
No single copy ever exceeds 1/N, so a moving edge resolves as a
staircase of ghosts behind a solid frame-time instance. The group
carries the element's own opacity, so a beat that moves and fades at
once fades its smear with it.
4. When every sampled transform collapses to the current one (element at
rest), the group is hidden so the element renders sharp.
Shutter window (AE semantics, per frame at time t, frame interval 1/fps):
shutterTime = shutterAngle / 360 / fps
windowStart = t + shutterPhase / 360 / fps
copy k = windowStart + k / N × shutterTime, k = 0..N (N+1 copies)
N counts sub-intervals, so both ends of the window carry a copy.
Defaults (shutterAngle 720, shutterPhase -360) integrate [t - 1/fps, t + 1/fps]:
a solid moving at 2000 px/s at 30 fps smears 133 px, centred on its position.
These defaults are measured, not assumed. On a 1920×1080 / 30 fps After Effects
export of translating text, the outermost duplicate on the trailing side sits
exactly at the previous frame's position and the outermost leading duplicate
exactly at the next frame's, with 8 evenly spaced duplicates in between on each
side (window = 2 frames = 720°, phase -360°, 16 sub-intervals). Reading the ghost
staircase across a stroke gives steps of 0.063 ± 0.002 of the sharp text
intensity — a flat 1/16 per duplicate, with no taper toward the window edges
(triangle weighting scores 1.6 dB worse against the same export). The sharp
instance on top is what keeps total ink above the unblurred frame's.
Requirements:
- Elements must be animated via `transform` (GSAP x/y/scale/rotation and the
3D and skew properties all qualify), not left/top. Transformed ancestors
are still not compensated.
- A 3D beat reads its perspective off the element's parent and applies it per
copy, so the vanishing point follows each copy's transform-origin rather
than the parent's perspective-origin. Those agree when the element is
centred in its perspective parent, which is how a 3D beat is authored.
- A copy drops the element's id but keeps its classes, because a class rule is
the only thing that can still style a copy's pseudo-elements. So a selector
written against a class matches the copies too: after attaching, address the
element by id or by reference, never by a class a copy also carries.
- A target is blurred once: a second attachMotionBlur() call naming the same
element leaves the first call's copies in place rather than stacking a
second set over them.
- Cost is N + 1 copies of the target's whole subtree, restyled on attach and
on resize and re-transformed every frame. Point it at the element that
moves, not at a container holding it.
- The host must render the timeline for the blur to update — HyperFrames
seeks every frame. A fresh paused timeline that is never seeked shows no
blur on its first frame.
- A declarative target is attached as soon as its composition's timeline is
registered, and the registry is also polled for about eight seconds after
load, so a composition that mounts asynchronously is picked up too. A
target added to the document after that window needs attachMotionBlur().
- GSAP must be loaded before this snippet executes.
API:
data-hf-motion-blur on the element that moves. Empty for defaults, else a
JSON OBJECT of the options below (double quotes on the keys). The
element's timeline is the one registered under the data-composition-id of
its nearest ancestor carrying that attribute, so a sub-composition's
targets follow that sub-composition.
Three things warn rather than rendering sharp in silence: a target no
composition ever registers a timeline for, a value that is not JSON, and
a value that parses but is not an object (null, a number, a string).
A second timeline registered under the same key also warns: the copies
keep following the first one.
attachMotionBlur(selector, timeline, options?) — the imperative form, for a
target built after the polling window. Call it AFTER defining all tweens,
so the timeline's final duration is known.
Either form blurs a given element once, whichever names it first.
Options:
shutterAngle — degrees of the frame interval the shutter is open
(default 720 = two frames, as measured off the AE
reference; 360 = one frame, 0 disables)
shutterPhase — degrees offset of the window start from the frame
time (default -360 = centred on the frame, one frame
back, like the AE reference)
samplesPerFrame — sub-intervals of the shutter window, so N+1 duplicates
each at 1/N opacity (default 16, max 64)
fps — composition frame rate (default: the data-fps of the
target's own composition root, else 30). Pass it
explicitly when rendering with an fps override
(`hyperframes render --fps`).
A key that is none of the four, or a value that is not a finite number, is
refused by name rather than read as defaults. Both halves matter: a
non-numeric angle hides the smear, and a non-numeric phase seeks every copy
to NaN, which parks them at the timeline's start.
-->
<script>
(function () {
/* SHUTTER_SNIPPET_START */
if (!window._hfMbUid) window._hfMbUid = 0;
if (!window._hfMbAttached) window._hfMbAttached = new WeakSet();
if (!window._hfMbSkipped) window._hfMbSkipped = new WeakSet();
if (!window._hfMbProxies) window._hfMbProxies = new WeakSet();
if (!window._hfMbTimeline) window._hfMbTimeline = new WeakMap();
if (!window._hfMbWarned) window._hfMbWarned = new WeakSet();
// The declarative form. A copy of a target must not carry it, or the copy is swept
// as a target of its own; the group's own marker is `data-hf-motion-blur-group`.
var SWEEP_ATTR = "data-hf-motion-blur";
// About eight seconds. Long enough for an asynchronously mounted sub-composition, short
// enough that a page with a genuine mistake in it reports before anyone stops watching.
var POLL_MS = 32;
var POLL_TICKS = 250;
var SWEEP_SELECTOR = "[" + SWEEP_ATTR + "]";
var OPTION_KEYS = ["shutterAngle", "shutterPhase", "samplesPerFrame", "fps"];
// A copy is a deep clone, so the attribute rides along on the root AND on every marked
// descendant. Any one left behind is swept as a target of its own, which clones again.
var COPY_STRIP = [SWEEP_ATTR, "data-composition-id", "data-fps"];
function stripAttrs(el) {
for (var a = 0; a < COPY_STRIP.length; a++) el.removeAttribute(COPY_STRIP[a]);
}
function stripCopyAttrs(root) {
stripAttrs(root);
var inner = root.querySelectorAll("[" + COPY_STRIP.join("],[") + "]");
for (var i = 0; i < inner.length; i++) stripAttrs(inner[i]);
}
// Shutter length is measured in frames, so the snippet must know the composition
// frame rate: explicit option, else the target's OWN root, else the first root in
// the document, else 30. The target's root is what makes two compositions on one
// page read their own data-fps instead of sharing the first one written.
function resolveFps(optionFps, target) {
var explicit = Number(optionFps);
if (explicit > 0) return explicit;
var scoped = target ? target.closest("[data-composition-id][data-fps]") : null;
var root = scoped || document.querySelector("[data-composition-id][data-fps]");
var rootFps = root ? Number(root.getAttribute("data-fps")) : 0;
if (rootFps > 0) return rootFps;
return 30;
}
function numOption(value, fallback) {
return value !== undefined ? Number(value) : fallback;
}
// A resolved `transform` is either "none", a 2D matrix(a,b,c,d,e,f) or a 3D
// matrix3d of 16. Splitting the numbers into the linear part and the
// translation is what lets the deadband below use a pixel tolerance on the
// one and an angle/ratio tolerance on the other.
function parseTransform(value) {
if (!value || value === "none") return { linear: [1, 0, 0, 1], translate: [0, 0, 0] };
var nums = value
.slice(value.indexOf("(") + 1, -1)
.split(",")
.map(parseFloat);
if (nums.length === 16) {
return {
linear: [
nums[0],
nums[1],
nums[2],
nums[4],
nums[5],
nums[6],
nums[8],
nums[9],
nums[10],
nums[3],
nums[7],
nums[11],
nums[15],
],
translate: [nums[12], nums[13], nums[14]],
};
}
return {
linear: [nums[0], nums[1], nums[2], nums[3]],
translate: [nums[4], nums[5], 0],
};
}
// Deadband: half a pixel of travel, a thousandth of a unit of linear change,
// about 0.06 degrees or 0.1% of scale. Below it the average is the source.
// The 2D and 3D forms have different arities, so a change of form counts as
// movement rather than being compared component by component.
function differs(a, b) {
if (a.linear.length !== b.linear.length) return true;
for (var i = 0; i < a.linear.length; i++) {
if (Math.abs(a.linear[i] - b.linear[i]) > 0.001) return true;
}
for (var j = 0; j < 3; j++) {
if (Math.abs(a.translate[j] - b.translate[j]) > 0.5) return true;
}
return false;
}
window.attachMotionBlur = function (selector, tl, opts) {
opts = opts || {};
var shutterAngle = numOption(opts.shutterAngle, 720);
var shutterPhase = numOption(opts.shutterPhase, -360);
var requestedSamples = numOption(opts.samplesPerFrame, 16);
var samples = Number.isFinite(requestedSamples)
? Math.max(2, Math.min(64, Math.round(requestedSamples)))
: 16;
var items = Array.isArray(selector) ? selector : [selector];
var targets = items.reduce(function (acc, s) {
if (typeof s === "string") {
document.querySelectorAll(s).forEach(function (el) {
acc.push(el);
});
} else if (s instanceof Element) {
acc.push(s);
}
return acc;
}, []);
// `samples` counts sub-intervals of the shutter window, so there is one copy at
// every interval boundary — both ends of the window included — and each carries
// 1/samples of the source. That is what the reference export shows: the outermost
// copy sits exactly one frame away from the frame time, not half a step short.
var copies = samples + 1;
var alpha = String(1 / samples);
// A copy is styled by nothing that selected the original: dropping the id also
// drops every `#id` rule that gave it size, colour and font. So each copy carries
// its own resolved style inline. Chrome returns "" for a computed style's cssText,
// so the declarations are enumerated rather than taken wholesale.
function resolvedStyles(source, out) {
var cs = getComputedStyle(source);
var text = "";
for (var i = 0; i < cs.length; i++) {
text += cs[i] + ":" + cs.getPropertyValue(cs[i]) + ";";
}
// A copy is a still of one instant. Left live, an inherited transition or
// keyframe animation would carry it somewhere the timeline never sampled.
out.push(text + "transition:none;animation:none;");
for (var c = 0; c < source.children.length; c++) {
resolvedStyles(source.children[c], out);
}
return out;
}
function paintResolvedStyles(copy, styles, cursor) {
copy.style.cssText = styles[cursor.i++];
for (var c = 0; c < copy.children.length; c++) {
paintResolvedStyles(copy.children[c], styles, cursor);
}
}
// Perspective applies to a parent's children only, and a copy is a grandchild.
// Sharing the parent's 3D context would need preserve-3d on the group, which
// mix-blend-mode flattens, so each copy carries the perspective itself. See the
// header for what that does to the vanishing point.
function parentPerspective(el) {
var value = el.parentNode ? getComputedStyle(el.parentNode).perspective : "none";
return value && value !== "none" ? "perspective(" + value + ") " : "";
}
function attachOne(el) {
var group = document.createElement("div");
group.setAttribute("data-hf-motion-blur-group", "hf-mb-" + window._hfMbUid++);
group.style.cssText =
"position:absolute;left:0;top:0;width:0;height:0;isolation:isolate;pointer-events:none;display:none;";
var clones = [];
for (var k = 0; k < copies; k++) {
var clone = el.cloneNode(true);
clone.removeAttribute("id");
stripCopyAttrs(clone);
clones.push(clone);
group.appendChild(clone);
}
var s = {
el: el,
group: group,
clones: clones,
samples: [],
perspective: "",
watch: null,
};
// Resolved styles and the parent's perspective are px once read, so a
// container-relative element needs them again when its box changes. Per frame
// that would walk the whole subtree once per copy; on resize it costs nothing
// on a fixed-size render and keeps a live preview honest.
function snapshot() {
var styles = resolvedStyles(el, []);
s.perspective = parentPerspective(el);
for (var c = 0; c < clones.length; c++) {
paintResolvedStyles(clones[c], styles, { i: 0 });
clones[c].style.position = "absolute";
clones[c].style.margin = "0";
clones[c].style.opacity = alpha;
clones[c].style.mixBlendMode = "plus-lighter";
}
}
// Snapshot BEFORE the group enters the page, so a throw in the style walk leaves
// nothing behind: the record the rollback needs does not exist until it returns.
snapshot();
el.parentNode.insertBefore(group, el);
if (typeof ResizeObserver === "function") {
s.watch = new ResizeObserver(snapshot);
s.watch.observe(el);
}
return s;
}
var state = [];
// Every element this call marked. A throw inside `attachOne` never returns its
// record, so `state` alone cannot give the marks back.
var claimed = [];
try {
for (var q = 0; q < targets.length; q++) {
var candidate = targets[q];
// One group per element, whether a second call names it or one call names it
// twice: a second set of copies over the first would double the ink. Marked
// BEFORE the copies go in: inserting them runs custom-element reactions
// synchronously, and one that registers a timeline re-enters the write trap.
if (window._hfMbAttached.has(candidate)) continue;
window._hfMbAttached.add(candidate);
window._hfMbTimeline.set(candidate, tl);
claimed.push(candidate);
state.push(attachOne(candidate));
}
} catch (e) {
unmark(claimed);
discard(state);
throw e;
}
// Nothing to blur means nothing to seek: a tracker here would run the whole sample
// loop every frame over an empty set.
if (state.length === 0) return;
// One seek loop drives the whole call, so the window is ONE length, and it belongs
// to the composition of an element this call actually attached.
var fps = resolveFps(opts.fps, state[0].el);
for (var d = 1; d < state.length; d++) {
if (resolveFps(opts.fps, state[d].el) === fps) continue;
var line = "one call cannot blur compositions at different frame rates; using ";
console.warn("[motion-blur] " + line + fps, state[d].el);
break;
}
function readTransform(el) {
var cs = getComputedStyle(el);
return {
css: cs.transform,
origin: cs.transformOrigin,
opacity: cs.opacity,
parsed: parseTransform(cs.transform),
};
}
var scheduled = false;
function applyShutter() {
var shutterTime = shutterAngle / 360 / fps;
if (!(shutterTime > 0)) {
state.forEach(function (s) {
s.group.style.display = "none";
});
return;
}
var t0 = tl.time();
var duration = tl.duration();
var windowStart = t0 + shutterPhase / 360 / fps;
// Sample the real trajectory: seek to each sub-frame time with events
// suppressed (so no tween callbacks, including this tracker's, fire),
// read the resolved transform, then restore the frame time.
try {
for (var k = 0; k < copies; k++) {
var tk = windowStart + (k / samples) * shutterTime;
tl.time(Math.min(duration, Math.max(0, tk)), true);
state.forEach(function (s) {
s.samples[k] = readTransform(s.el);
});
}
} finally {
tl.time(t0, true);
}
state.forEach(function (s) {
var current = readTransform(s.el);
var moved = false;
for (var i = 0; i < copies; i++) {
if (differs(s.samples[i].parsed, current.parsed)) {
moved = true;
break;
}
}
if (!moved) {
s.group.style.display = "none";
return;
}
// The copies are absolutely positioned inside the group, and the group sits at
// the origin of the element's own containing block, so the element's offset box
// places them. Re-read every frame: a beat is free to move the box itself.
var left = s.el.offsetLeft + "px";
var top = s.el.offsetTop + "px";
var width = s.el.offsetWidth + "px";
var height = s.el.offsetHeight + "px";
for (var j = 0; j < copies; j++) {
var clone = s.clones[j];
clone.style.left = left;
clone.style.top = top;
clone.style.width = width;
clone.style.height = height;
clone.style.transformOrigin = s.samples[j].origin;
clone.style.transform =
s.perspective + (s.samples[j].css === "none" ? "" : s.samples[j].css);
}
// The copies are the element's own ink, so they have to carry its opacity too:
// a beat that moves and fades at once would otherwise leave a full-strength
// smear behind a vanishing element. It rides the group, not the copies, whose
// own opacity is the 1/N shutter weight.
s.group.style.opacity = current.opacity;
s.group.style.display = "";
});
}
// tl.eventCallback("onUpdate") is unavailable under the HyperFrames runtime proxy, so a
// tracker tween's onUpdate fires on every seek. Sampling is deferred to a microtask:
// seeking GSAP from inside its own render reads stale tween state, while the microtask
// runs after the seek returns and before the frame is captured or painted.
var _proxy = { t: 0 };
try {
tl.to(
_proxy,
{
t: 1,
duration: Math.max(tl.duration(), 0.1),
ease: "none",
onUpdate: function () {
if (scheduled) return;
scheduled = true;
Promise.resolve().then(function () {
scheduled = false;
applyShutter();
});
},
},
0,
);
} catch (e) {
unmark(claimed);
discard(state);
throw e;
}
};
// An empty attribute means defaults; anything else is a JSON options object. A value that
// is not an options object is skipped loudly rather than read as defaults, because a typo
// in it would otherwise render as a silently unblurred element.
function sweepOptions(el) {
var text = (el.getAttribute(SWEEP_ATTR) || "").trim();
if (text === "") return {};
var parsed = null;
try {
parsed = JSON.parse(text);
} catch (e) {
return refuse(el, "is not JSON: " + text);
}
// `null`, a number and a string all parse. None of them is a set of options.
if (parsed === null || typeof parsed !== "object" || Array.isArray(parsed)) {
return refuse(el, "is not a JSON object: " + text);
}
var bad = unknownKey(parsed);
if (bad !== null) return refuse(el, "names no such option: " + bad);
var wrong = nonNumber(parsed);
if (wrong !== null) return refuse(el, "needs a number for " + wrong);
return parsed;
}
// A misspelled key is the likeliest typo in a JSON object, and reading it as
// defaults is the silent no-op this attribute exists to remove.
function unknownKey(parsed) {
for (var key in parsed) {
if (OPTION_KEYS.indexOf(key) < 0) return key;
}
return null;
}
// A key with a value the shutter cannot use is the same silent no-op as a key it does
// not know: a non-numeric angle hides the smear, a non-numeric phase parks every copy
// at the timeline's start. JSON gives real numbers, so anything else is a typo.
function nonNumber(parsed) {
for (var i = 0; i < OPTION_KEYS.length; i++) {
var held = parsed[OPTION_KEYS[i]];
if (held === undefined) continue;
if (typeof held !== "number" || !Number.isFinite(held)) return OPTION_KEYS[i];
}
return null;
}
// A throw partway through an attach leaves an element blurred by nothing, so the copies
// go too, along with the observer that would otherwise keep walking a detached subtree
// on every resize of an element nothing blurs.
// The mark is what a second trigger reads, so it is the first thing to give back.
function unmark(els) {
for (var i = 0; i < els.length; i++) {
window._hfMbAttached.delete(els[i]);
window._hfMbTimeline.delete(els[i]);
}
}
function discard(state) {
for (var i = 0; i < state.length; i++) {
var s = state[i];
if (s.watch !== null) s.watch.disconnect();
if (s.group.parentNode !== null) s.group.parentNode.removeChild(s.group);
}
}
function refuse(el, why) {
window._hfMbSkipped.add(el);
console.warn("[motion-blur] " + SWEEP_ATTR + " " + why, el);
return null;
}
// An element belongs to the nearest composition above it, so a sub-composition's
// targets follow that sub-composition's own timeline.
function compositionOf(el) {
var host = el.closest("[data-composition-id]");
return host ? host.getAttribute("data-composition-id") : null;
}
// Anything else registered on the container is somebody else's bookkeeping: a
// timeline is the thing this snippet can seek and extend.
function isTimeline(value) {
return !!value && typeof value.time === "function" && typeof value.to === "function";
}
// Attached, or skipped as unparseable. Both are terminal, so a later trigger must not
// report or attach the element twice.
function handled(el) {
return window._hfMbAttached.has(el) || window._hfMbSkipped.has(el);
}
// The timeline registered for an element's own composition, or null. Read from the
// registry rather than taken from whoever wrote it, so every trigger agrees.
function timelineFor(el) {
var id = compositionOf(el);
if (id === null) return null;
var registry = window.__timelines;
var found = registry ? registry[id] : null;
return isTimeline(found) ? found : null;
}
// A marked element whose parent is not an Element has no target above it: the parent
// of a root is the Document, which has no `closest`.
function nestedInTarget(el) {
var parent = el.parentNode;
if (parent === null || typeof parent.closest !== "function") return false;
return parent.closest(SWEEP_SELECTOR) !== null;
}
// One owner of attaching one element. Answers whether the element still needs a trigger.
function claim(el) {
var tl = timelineFor(el);
if (handled(el)) {
var owner = window._hfMbTimeline.get(el);
// Re-registering a key with a DIFFERENT timeline leaves the copies driven by the one
// nobody seeks, which renders sharp and says nothing. Report it rather than guess.
var changed = tl !== null && owner !== undefined && owner !== tl;
if (changed && !window._hfMbWarned.has(el)) {
window._hfMbWarned.add(el);
console.warn(
"[motion-blur] a second timeline registered for this element; the copies still follow the first",
el,
);
}
return true;
}
// Its group would be inserted into the live outer element, so the outer target's
// style replay would walk a subtree its cursor no longer matches.
if (nestedInTarget(el)) {
refuse(el, "cannot blur a target inside another target; blur one of them");
return true;
}
var opts = sweepOptions(el);
if (opts === null) return true;
if (tl === null) return false;
window.attachMotionBlur(el, tl, opts);
return true;
}
// Returns the elements still waiting for their composition to register a timeline.
// The trap below runs this INSIDE the author's registration statement, so one bad
// target must not abort the rest of their composition script.
function sweep() {
var pending = [];
document.querySelectorAll(SWEEP_SELECTOR).forEach(function (el) {
try {
if (!claim(el)) pending.push(el);
} catch (e) {
window._hfMbSkipped.add(el);
console.warn("[motion-blur] could not blur this element: " + e, el);
}
});
return pending;
}
// Registering a timeline is the moment the author has finished building it, so trapping
// the write attaches before the renderer captures a frame. A microtask would be too
// early: it runs before the next `<script>` tag.
function registrationProxy(container) {
var proxy = new Proxy(container, {
set: function (target, key, value) {
target[key] = value;
if (typeof key === "string") sweep();
return true;
},
});
window._hfMbProxies.add(proxy);
return proxy;
}
function hookRegistration() {
var held = window.__timelines;
Object.defineProperty(window, "__timelines", {
configurable: true,
enumerable: true,
get: function () {
return held;
},
set: function (value) {
var container = !!value && typeof value === "object";
var fresh = container && !window._hfMbProxies.has(value);
held = fresh ? registrationProxy(value) : value;
if (fresh) sweep();
},
});
// A container assigned before this snippet ran is replayed through the setter, so a
// timeline already registered on it is swept too.
if (held !== undefined) window.__timelines = held;
}
// The trap above cannot see every registration: the runtime owns the registry object,
// wraps it per sub-composition, and mounts nested compositions asynchronously. So the
// registry is also POLLED, which needs no cooperation from whoever writes it.
function poll(ticks) {
var pending = sweep();
// The WHOLE budget, every time. Stopping the first tick nothing is pending kills the
// poll one tick into an ordinary page, which is before a late composition arrives.
if (ticks > 0) {
setTimeout(function () {
poll(ticks - 1);
}, POLL_MS);
return;
}
if (pending.length === 0) return;
var count = pending.length + " element(s) carry " + SWEEP_ATTR;
console.warn(
"[motion-blur] " + count + ", and no composition registered a timeline for them",
pending,
);
}
if (!window._hfMbHooked) {
window._hfMbHooked = true;
hookRegistration();
poll(POLL_TICKS);
}
/* SHUTTER_SNIPPET_END */
})();
</script>
<!--
Timeline integration example:
<div id="root" data-composition-id="my-composition" data-duration="6" data-fps="30">
<div id="my-box" data-hf-motion-blur></div>
<div id="my-title" data-hf-motion-blur='{"shutterAngle": 360}'></div>
</div>
const tl = gsap.timeline({ paused: true });
tl.fromTo("#my-box", { x: -100 }, { x: 1700, duration: 1.2, ease: "power3.inOut" }, 0.5);
// Extend to data-duration so seeks past the last tween reach the blur callback.
tl.set(document.body, {}, DATA_DURATION);
// The registration below is what attaches both targets, in that order.
window.__timelines = window.__timelines || {};
window.__timelines["my-composition"] = tl;
// Only for an element created later than the polling window above:
attachMotionBlur("#late-box", tl, { shutterAngle: 720, samplesPerFrame: 16 });
-->
Usage
Paste the snippet into your composition, then putdata-hf-motion-blur on any element your timeline animates. Nothing else to call, and no order to get wrong.
<div id="root" data-composition-id="my-composition" data-duration="4" data-fps="30">
<!-- Defaults: a two-frame shutter at 16 sub-intervals. -->
<div id="my-box" data-hf-motion-blur></div>
<!-- Or a JSON object of options. An unknown name, or a value that is not a
number, is refused by name rather than read as defaults. -->
<div id="my-title" data-hf-motion-blur='{"shutterAngle": 360, "samplesPerFrame": 8}'></div>
</div>
[data-composition-id] ancestor names the key, and the snippet reads window.__timelines[key]. An element whose composition never registers a timeline says so in the console instead of rendering sharp in silence.
attachMotionBlur() stays available for a target built after the page has settled. It has an ordering contract the attribute does not: call it after every tween is defined, so the timeline’s final duration is known.
<!-- Extend the timeline to data-duration before calling attachMotionBlur -->
tl.set(document.body, {}, DATA_DURATION);
attachMotionBlur("#late-box", tl, {
shutterAngle: 720, // degrees of the frame interval the shutter is open
samplesPerFrame: 16, // sub-intervals of the window; 17 duplicates at 1/16 each
});
How it works
This is the After Effects layer Motion Blur model — temporal supersampling — synthesised inside the page, because the renderer captures one instant per frame with no shutter.- Sample the trajectory — after every timeline seek, the snippet seeks the timeline (events suppressed) to one sub-frame time per window boundary, reads each target’s resolved
transformmatrix, then restores the frame time. Reading the resolved matrix rather than a list of named GSAP properties is what makes every driver work: translation, scale, rotation, 3D rotation and skew all arrive in the same one value. The window isshutterAngle / 360 / fpsseconds long and startsshutterPhase / 360 / fpsseconds from the frame time;samplesPerFramedivides it into that many sub-intervals, so there aresamplesPerFrame + 1samples and both ends of the window carry one. - Add the duplicates — each target gets a group of DOM copies of itself, one per sample, each carrying that sample’s transform at
1/Nopacity withmix-blend-mode: plus-lighter. plus-lighter adds premultiplied colour, so N + 1 copies at 1/N sum to the average, which is the shutter integral. The group setsisolation: isolate, and that is load-bearing rather than tidy: without a transparent backdrop of its own the first copy would add onto the page and blow a light background out to white. A solid translating at speedvbecomes a box smear of lengthv × shutterTime, centred on its instantaneous position — no one-sided trail and no Gaussian halo. No single duplicate exceeds1/N, so the tails step down evenly; where the element travels less than its own size the copies overlap past 1 and clamp, leaving the middle solid rather than a plateau. - Composite the sharp instance on top — the element is then drawn over the smear at full opacity, so the position the frame is actually at stays crisp and only the duplicates are faint. This is the single largest difference from a plain average, and the one that makes the result read like After Effects rather than like a long exposure.
- Rest is sharp — when every sampled transform collapses onto the current one, the group is hidden and the element renders unblurred. The group also carries the element’s own opacity, so a beat that moves and fades at once fades its smear with it.
transform drives the smear on its own: a beat that only scales or only turns blurs exactly as a beat that only translates does. A 3D beat reads its perspective off the element’s parent and applies it per copy, because mix-blend-mode flattens preserve-3d and a copy therefore cannot inherit the parent’s 3D context. Transformed ancestors are still not compensated. A copy drops the element’s id but keeps its classes, so after attaching, address the element by id or by reference rather than by a class the copies also carry.
Options
| Option | Default | Description |
|---|---|---|
shutterAngle | 720 | Degrees of the frame interval the shutter is open. 720 spans two frames, the width measured off the After Effects reference; 360 spans one; 0 disables the blur |
shutterPhase | -360 | Degrees offset of the window start from the frame time. -360 starts the window one frame back, which with a 720° angle centres it on the frame |
samplesPerFrame | 16 | Sub-intervals of the shutter window (2–64), so N + 1 duplicates at 1/N opacity each. More sub-intervals close the gaps between duplicates at high speed, at the cost of one more copy of the element per frame |
fps | root data-fps, else 30 | Composition frame rate. Pass it explicitly when rendering with an fps override (hyperframes render --fps) |
effect motion-blur shutter after-effects velocity animation.