guillermolg00/morphicons

morphicons.com — live playground

Universal morphing for stroke-based icons (Lucide, Tab ler, Heroicons, Iconoir, or your own paths): any icon morphs into any other with spring physics. Rotations are never declared by hand — they emerge from the math (2D Procrustes + polar interpolation). Zero runtime dependencies.

import { MorphIcon } from "morphicons/react";
import { Menu, X } from "lucide"; // data, not components

<button onClick={() => setOpen(o => !o)} aria-expanded={open}>
  <MorphIcon icon={open ? X : Menu} />
</button>

That's the whole thing. No wrappers, no AnimatePresence, no keys, no from/to pairs, no configuration. State lives outside; the animation is an implementation detail the component picks up when the prop changes.

The usual icon morphs either interpolate raw coordinates (shapes shrink and shear in transit) or require hand-declaring "rotation groups" per icon pair. morphicons solves the optimal similarity between shapes in closed form and interpolates it in its natural space: if a pair is congruent under rotation, it rotates; if not, it morphs in the aligned frame. arrow-right → arrow-down yields θ = 90° on its own, without anyone declaring it.

bun add morphicons        # or npm install / pnpm add

ESM only. react (>= 18), vue (>= 3.3), svelte (>= 5) and react-native (>= 0.71) + react-native-svg (>= 14) are optional peers — only needed for morphicons/react, morphicons/vue, morphicons/svelte and morphicons/react-native. morphicons/element (the <morph-icon> custom element) and morphicons/astro (its SSR shell) need no peer at all — Astro compiles the shell itself, like every .astro component shipped as source.

Icons come from a data package, not a component package. morphicons consumes icon data (an IconNode or a raw d string). For Lucide that means the vanilla lucide package: import { Menu, X } from "lucide" gives you IconNodes, which is why every snippet below says "data, not components". The framework packages (lucide-react, lucide-vue-next, @lucide/svelte, lucide-react-native) export components, and MorphIcon can't consume those. If your app already renders static icons with one of them, keep it: the data and component packages coexist by design and both tree-shake cleanly, so you only pay for the icons you import. Just keep their versions aligned, so the icons you morph match the ones you render statically.

import { MorphIcon, type MorphHandle } from "morphicons/react";

// 1. Uncontrolled (90% of uses): change the prop and morphicons animates
<MorphIcon icon={open ? X : Menu} spring="snappy" />

// 2. Controlled (gestures, scroll): explicit progress, no spring
<MorphIcon from={Menu} to={X} progress={dragProgress} />

// 3. Imperative (sequences)
const ref = useRef<MorphHandle>(null);
<MorphIcon ref={ref} icon={Menu} />
ref.current?.morphTo(Check); // animates
ref.current?.set(X);         // jumps without animating

Drop-in replacement for lucide-react (the props surface: the icon prop takes Lucide data, not lucide-react components, see Install): size, strokeWidth, absoluteStrokeWidth, color, className and the rest of the <svg> props pass straight through. Correct accessibility by default: aria-hidden unless you pass label (→ role="img" + <title>). Clean SSR: the server emits the exact static SVG (zero flash, zero layout shift); the runtime is born on hydration. Morphs play regardless of the OS reduce-motion setting by default; opt into honoring it with reducedMotion="user" (see Reduced motion).

<script setup lang="ts">
import { ref } from "vue";
import { MorphIcon, type MorphHandle } from "morphicons/vue";
import { Menu, X, Check } from "lucide"; // data, not components

const open = ref(false);
const morph = ref<MorphHandle | null>(null);
// morph.value?.morphTo(Check)  → animates
// morph.value?.set(X)          → jumps without animating
</script>

<template>
  <!-- 1. Uncontrolled (90% of uses): change the prop and morphicons animates -->
  <MorphIcon :icon="open ? X : Menu" spring="snappy" />

  <!-- 2. Controlled (gestures, scroll): explicit progress, no spring -->
  <MorphIcon :from="Menu" :to="X" :progress="dragProgress" />

  <!-- 3. Imperative (sequences): template ref → morphTo / set -->
  <MorphIcon ref="morph" :icon="Menu" />
</template>

Same surface as the React binding: presentation props (size, strokeWidth, absoluteStrokeWidth, color; class, style and the rest of the <svg> attrs fall through), the same accessibility defaults (aria-hidden unless you pass label) and the same clean SSR — works with Nuxt out of the box: the server emits the exact static SVG and the runtime is born on hydration. The binding is a plain render function; no SFC compiler or JSX involved.

<script lang="ts">
  import { MorphIcon, type MorphHandle } from "morphicons/svelte";
  import { Menu, X, Check } from "lucide"; // data, not components

  let open = $state(false);
  let morph = $state<MorphHandle>();
  // morph?.morphTo(Check)  → animates
  // morph?.set(X)          → jumps without animating
</script>

<!-- 1. Uncontrolled (90% of uses): change the prop and morphicons animates -->
<MorphIcon icon={open ? X : Menu} spring="snappy" />

<!-- 2. Controlled (gestures, scroll): explicit progress, no spring -->
<MorphIcon from={Menu} to={X} progress={dragProgress} />

<!-- 3. Imperative (sequences): bind:this → morphTo / set -->
<MorphIcon bind:this={morph} icon={Menu} />

Same surface again: presentation props, class/style/rest-attr fall-through — fully typed via svelte/elements, so SVG attrs, events and ARIA autocomplete and typos fail the build, like lucide-svelte — the same accessibility defaults and the same clean SSR — works with SvelteKit out of the box. Svelte 5 (runes); the component ships as .svelte source and your bundler compiles it via the svelte export condition, like every Svelte library. If you followed Lucide's official Svelte guide you already have @lucide/svelte: keep it for your static icons; MorphIcon takes the data exports from the vanilla lucide package (see Install).

import { MorphIcon, type MorphHandle } from "morphicons/react-native";
import { Menu, X, Check } from "lucide"; // data, not components

// 1. Uncontrolled (90% of uses): change the prop and morphicons animates
<MorphIcon icon={open ? X : Menu} spring="snappy" />

// 2. Controlled (gestures, scroll): explicit progress, no spring
<MorphIcon from={Menu} to={X} progress={dragProgress} />

// 3. Imperative (sequences)
const ref = useRef<MorphHandle>(null);
<MorphIcon ref={ref} icon={Menu} />
ref.current?.morphTo(Check); // animates
ref.current?.set(X);         // jumps without animating

The DOM-free core beyond the browser: the dom driver is reused verbatim as the engine (React Native has a global requestAnimationFrame, and PathEl is structural), so the whole platform difference is a shim that forwards the per-frame d write to Path.setNativeProps of react-native-svg — outside the React render, exactly like the web mutation. On the New Architecture that write is not the last word, because every Fabric commit rebuilds the native path from the props React declares: the binding therefore publishes the driver's live d on each render, and re-applies it in a layout effect when a frame landed in between, so an unrelated re-render (a color or selection change, at rest or mid-flight) can no longer restore the icon the component mounted with. Same surface as the React binding (size, strokeWidth, absoluteStrokeWidth, color, plus the native Svg props: testID, touch handlers…), same accessibility defaults (aria-hidden unless you pass label → role="img" + aria-label). With reducedMotion="user", the OS setting comes from AccessibilityInfo (best-effort: the query is async; a reduceMotionChanged subscription keeps it exact from then on). Requires Metro with package exports resolution — default since React Native 0.79; on older versions enable unstable_enablePackageExports.

---
import MorphIcon from "morphicons/astro";
import { Menu } from "lucide"; // data, not components
---

<MorphIcon icon={Menu} label="Menu" id="menu-icon" />

<script>
  import { X } from "lucide";
  import type { MorphIconElement } from "morphicons/element";
  // Post-load control happens ON the element (it upgrades to <morph-icon>):
  const el = document.querySelector<MorphIconElement>("#menu-icon")!;
  el.morphTo(X);             // 3. imperative: morphTo / set, like a ref
  // el.icon = X;            // 1. uncontrolled: assigning the property animates
  // el.from = Menu; el.to = X; el.progress = 0.5;  // 2. controlled: frozen seek
</script>

The server emits the exact static SVG with the pure core (zero flash, zero layout shift, works with static output and any SSR adapter) and hydration is just custom-element upgrade: the shell defines <morph-icon> once per page and the element adopts the server markup verbatim — zero d writes during upgrade, pinned by instrumented tests. No framework runtime ships — the only client bytes are morphicons/element. A controlled pair of d strings survives upgrade as attributes, so assigning progress from a script scrubs the server-frozen pair; IconNode pairs are server-only (attributes can't carry them) — control those by assigning the element's properties. Same presentation props as every binding (size, strokeWidth, absoluteStrokeWidth, color, label); rest attrs (id, class, data-*…) land on the <morph-icon> element, because grabbing it from a client script is the interaction model here. React/Vue/Svelte islands keep working for icons that live inside one — this entry is for the pages that don't need an island at all.

The Astro shell is a thin SSR layer over morphicons/element, which stands on its own in plain HTML, HTMX, Rails, or any server-rendered stack:

<script type="module">
  import { defineMorphIcon } from "morphicons/element";
  defineMorphIcon(); // idempotent; custom tag: defineMorphIcon("my-icon")
</script>

<!-- attributes take d strings; properties take the full IconInput -->
<morph-icon icon="M4 6h16M4 12h16M4 18h16" label="Menu"></morph-icon>

The element is the fifth binding, not a second implementation: it wires attributes and properties to the same shared controller, passes the same mirrored mount suite, and exposes the same three modes — icon (uncontrolled), from/to/progress (controlled), morphTo/set methods (imperative, the element IS the handle). Attributes carry strings (d paths, preset names); properties accept everything (IconNodes, custom springs). If the element already contains a server-rendered <svg><path>, it adopts it verbatim and treats its d as the at-rest icon — SSR bytes are never rewritten at rest.

The five components share one contract, pinned by mirrored client-mount tests:

• Lazy driver. Mounting without any icon is fine — SSR emits <path d=""> and the driver is born with the FIRST icon that shows up, whether a late icon prop (data that loads async), a late from/to pair, or an imperative set/morphTo. The first icon paints without animating; morphTo before the driver exists behaves as set (there is nothing to fly from).
• Controlled wins. While from AND to are both present, the pair owns the path: icon changes are ignored, no spring fires. Drop the pair and the current icon takes over (animated). Mixing the modes is not an error — the precedence is just explicit.
• Clean re-entry. Any exit from controlled mode (an imperative call or an icon takeover) invalidates the frozen pair, so returning to the same from/to re-bases on from and renders exactly like a clean mount at that progress.

Icon morphs are small, short, communicative micro-transitions: the kind of motion the reduce-motion guidance considers generally acceptable, unlike parallax or full-screen movement. Auto-degrading them made the library look broken to every user with the OS setting on, so since 1.4.2 they play by default and the policy is an explicit prop (the same default Motion, formerly Framer Motion, ships):

<MorphIcon icon={open ? X : Menu} reducedMotion="user" />

• "never" (default): morphs always animate; the OS setting is ignored.
• "user": while the OS reduce-motion setting is on, every morphTo degrades to an instant swap (the pre-1.4.2 behavior). The web bindings read prefers-reduced-motion; React Native reads AccessibilityInfo, and only touches the accessibility bridge when some instance opts in.
• "always": every morphTo jumps without animating (tests, screenshots).

The prop is live: changing it governs the next morph. The vanilla driver takes the same policy as an option, createMorph(el, icon, { reducedMotion: "user" }), and exposes it as a mutable morph.reducedMotion. Apps that must strictly follow the user's system preference should pass reducedMotion="user" across the board.

import { createMorph } from "morphicons/dom";

const m = createMorph(pathEl, Menu); // pathEl: any object with setAttribute
m.morphTo(X, "snappy");              // preset or { stiffness, damping }
m.set(Check);                        // jump without animating (canonical d)
m.seek(X, 0.4);                      // morph frozen at t — scrubbing, gestures
m.progress = 0.4;                    // sugar: seek on the active target
m.reducedMotion = "user";            // live policy: "never" (default) | "user" | "always"
m.destroy();

Truly interruptible: a morphTo mid-flight re-plans from the current intermediate shape while preserving the spring's velocity — click spam never jumps. A single global requestAnimationFrame drives all instances.

import { resampleIcon, buildPlan, interpPolar, allocOutputs, serialize } from "morphicons";

const plan = buildPlan(resampleIcon(Menu), resampleIcon(X)); // cacheable
const out = allocOutputs(plan);
interpPolar(plan, 0.5, out);                                 // t ∈ [0, 1]
const d = serialize(out, plan.items.map(it => it.closed));   // `d` attribute

Accepts IconNode (Lucide's data format, structurally typed — Lucide is neither a dependency nor a peer) or a raw d attribute. Tabler, Heroicons, Iconoir and custom paths work the same.

The core has exactly two structural contracts: IconInput on the way in (an IconNode or a d string) and PathEl on the way out (anything with setAttribute — the same contract that lets the React Native binding drive setNativeProps). A format adapter converts a foreign icon representation into one of those two contracts. All of them live in one opt-in entry, so the unit of payment is the import: core, driver and bindings never grow for a capability you didn't ask for, and inside the entry each adapter is an independent module — a named import tree-shakes to just that adapter, pinned by per-export size gates in CI. Adapters are per format, never per vendor: there is no /iconify and never will be.

For icons that exist as markup rather than data: an Iconify body from @iconify-json/* (what UnoCSS presetIcons and the Tailwind icon plugins consume), a full <svg> copied from a website, a <path> from the shadcn registry. One call turns it into an IconInput any entry accepts:

import { svgToIcon } from "morphicons/adapters";

// Parse ONCE, at module scope — reuse the reference so the plan cache holds.
const MENU = svgToIcon(menuBody); // '<path stroke="currentColor" d="M4 5h16…"/>'
const X = svgToIcon(xBody);       // works in createMorph and all four bindings

Everything else happens under the hood: non-rendered containers (<defs>, <mask>, <clipPath>, <symbol>) are stripped, the morphability scan rejects what can't honestly morph (a fill-drawn icon with no stroke — Material Symbols style — a transform, an element outside the seven stroke primitives) with a clear error at parse time, and when the markup carries a viewBox the geometry is re-gridded onto 24 via fitIcon automatically, so off-grid collections just work.

For apps that render icons as CSS masks — UnoCSS presetIcons, @iconify/tailwind, @egoist/tailwindcss-icons (i-lucide-*, icon-[lucide-x]): the icon is a mask-image data URI on a <span>/<div>, with no <path> in the DOM to write to. maskTarget creates a hidden pair of referenced <svg><mask><path> buffers once, points the element's mask at them, and hands the driver a PathEl whose d writes land on that real geometry — so the element morphs in place, keeping your size-*/text-* classes. Two engine traps shaped this mechanism (see docs/adr/0002): a data-URI mask swapped per frame never decodes in time in Chromium (blank flight), and WebKit doesn't repaint when a referenced mask's content mutates (frozen flight) — hence the pair: each write flips mask-image between the two buffers, which forces re-resolution everywhere.

import { maskTarget } from "morphicons/adapters";
import { createMorph } from "morphicons/dom";

const target = maskTarget(spanEl);     // spanEl: class="size-5 text-current"
const m = createMorph(target, MENU);
m.morphTo(X, "snappy");                // full Morph handle: set/seek/progress/destroy
// on unmount: m.destroy(); target.dispose();

It is createMorph — only the write target differs — so springs, reduced motion and mid-flight interruption behave identically. On setup it sets background-color: currentColor, so a bare <span> works with no icon class ({ paint: false } keeps an existing bg-*; strokeWidth and viewBox are options). The returned target adds dispose() — call it after morph.destroy() when the element unmounts, so the hidden mask nodes go with it.

Cost. A live inline <path> gets the cheapest possible write; a mask element additionally makes the browser re-rasterize the mask and repaint the masked box every frame (main thread, no compositor-only path). Great for toggles and short lists (menu↔close, chevrons, play↔pause); for icon-heavy views, the inline-SVG bindings remain the leanest option — same styling, one less indirection.

For surfaces where the DOM never enters: the icon stops being a document node and becomes pixels you own. The platform does the heavy lifting — Path2D parses d strings natively — so every frame is one stroke(new Path2D(d)) on whatever you hand it: an HTMLCanvasElement, an OffscreenCanvas, or a bare 2D context (what a worker, or a host that already owns the context, has in hand).

import { canvasTarget } from "morphicons/adapters";
import { createMorph } from "morphicons/dom";

const m = createMorph(canvasTarget(canvas), MENU);
m.morphTo(X, "snappy");
// the canvas is a live texture: gl.texImage2D(gl.TEXTURE_2D, ..., canvas)

Once the icon is pixels, everything the SVG model can't reach opens up:

• Canvas-first apps with no DOM to put an icon in: Konva, Fabric.js, PixiJS, games, chart draws, tldraw-style editors.
• OffscreenCanvas + workers: the whole morph runs off the main thread.
• Native export: toBlob() for PNG sprites, captureStream() + MediaRecorder for the morph as a video asset.
• An animated favicon: draw, toDataURL(), swap the <link>.
• HUDs over video or fullscreen surfaces: same buffer as the player or the game, no DOM layer on top.
• WebGL textures: the canvas is a ready texImage2D source — shader effects over live icons are one branch of this list, not its headline.
• Down the road, the same gesture (PathEl → Path2D) is PathEl → SkPath: a react-native-skia / CanvasKit target would consume the identical contract.

It is createMorph — springs, mid-flight interruption and reduced motion behave identically. The target maps the icon grid onto whatever backing store it finds, per write: min-side scale, centered (the canvas equivalent of preserveAspectRatio="xMidYMid meet"), so non-square canvases letterbox and devicePixelRatio is simply whatever you sized the backing store to. Color cascades: the color option wins; without it, a CSS-styled canvas contributes its computed color (the moral equivalent of currentColor, which canvas lacks — read lazily, once); where CSS can't reach (OffscreenCanvas, workers) the context's current strokeStyle stands, which is the idiomatic canvas contract. strokeWidth and viewBox mirror the other adapters; clear: false skips the per-frame clear for manual compositing and trails.

Hosts with their own render loop. Don't hand a Konva/Chart.js/game context to a morph — the driver writes on its own rAF and the two clocks fight over the frame. Give the icon a small dedicated canvas and composite it into the scene (drawImage, Konva.Image, a texture upload), with onWrite as the dirty signal:

const icon = new OffscreenCanvas(64, 64);
const target = canvasTarget(icon, {
  color: "#e8eaed",                      // CSS can't reach an OffscreenCanvas
  onWrite: () => { textureDirty = true } // re-upload / blit on the host's next frame
});
const m = createMorph(target, PLAY);

Cost. The cheapest target of the three: no layout, no DOM paint — one native path parse and one stroke rasterization per frame, on a canvas that is typically small. 0.48 KB gzip standalone, the lightest adapter yet.

morphicons has no per-library adapters — any icon set that meets these requirements works out of the box:

1. Stroke-drawn icons. The geometry must be the stroked centerline (fill="none", color via stroke). The whole pipeline — resampling, correspondence, in-flight polylines — assumes strokes; filled or outlined-fill glyphs (Material Symbols, Bootstrap Icons, Remix Icon, Phosphor, Heroicons solid) parse fine but won't read correctly in transit.

2. Geometry available as data. Either a raw d attribute per icon, or a [tag, attrs][] node list (Lucide's IconNode shape) using only path, line, circle, ellipse, rect, polyline, polygon. No <g> wrappers and no transform attributes — coordinates must be literal. Any other tag throws a clear error.

3. A shared coordinate space per pair. Both endpoints of a morph must live on the same grid. Lucide, Tabler, Heroicons and Iconoir all draw on 24×24 — that's why cross-library morphs just work. For a pack on another canvas (Heroicons solid on 20, Carbon on 32, Teenyicons on 15), re-grid it once with fitIcon:

   import { fitIcon } from "morphicons";
   
   const search = fitIcon(carbonSearch, 32);   // 32×32 → the 24 grid
   const bell   = fitIcon(teenySmallBell, 15); // also "0 0 15 15" or [0, 0, 15, 15]

   It returns a plain d, so it goes anywhere an icon is accepted. Call it at module scope, not per render. Skipping it doesn't throw — Procrustes is similarity-invariant, so the grid mismatch never reads as false rotation; it lands in σ as an unwanted zoom, and the target draws outside the canvas. The React, Vue and Svelte bindings default to viewBox="0 0 24 24", overridable via props.

4. Uniform stroke. Cosmetic rather than structural: a consistent stroke width and round caps/joins make in-flight shapes look native. These live on the <svg>, not in morphicons.

Beyond the libraries in the playground, sets known to qualify on the 24×24 grid — no fitting needed: Heroicons (outline style), Iconoir (1.5px stroke), Akar Icons, Untitled UI and Hugeicons (stroke style). Off-grid packs work through fitIcon: Teenyicons (15×15), Carbon (32×32), Heroicons solid (20×20).

This also covers the shadcn registry, which re-publishes 200+ icon libraries as React components with an inline <path d> — pass that d straight in, adding fitIcon when the pack's viewBox isn't 24.

┌─────────────────────────────────────────────┐
│  bindings   react · vue · svelte · rn      │  thin: ref + createMorph
│             · element (+ astro SSR shell)   │
├─────────────────────────────────────────────┤
│  drivers    dom (setAttribute + rAF)        │  singleton scheduler
├─────────────────────────────────────────────┤
│  core       parse → normalize → resample    │  pure functions,
│             → match → align → plan          │  NO DOM, produce
│             → interpolate → serialize       │  `d` strings and numbers
│             + spring                        │
└─────────────────────────────────────────────┘

The hard rule: the core never touches the DOM. Pure functions consume icon data and produce d strings. Direct consequence: the React Native binding over react-native-svg is just another adapter, not a rewrite — it reuses the dom driver verbatim through a setNativeProps shim. One package with subpath exports (. core, ./dom, ./react, ./react-native, ./vue, ./svelte, ./element, ./astro), ESM only, sideEffects: false.

The plan is the central artifact:

icon A ─┐
        ├→ normalize → resample → match → align → PLAN ─→ interpolate(t) → serialize → d
icon B ─┘                                           ↑
                                        cacheable, serializable

plan() accepts any list of sampled subpaths — not just canonical icons — which is what makes interruptions free: the currently rendered buffers are a valid morph source.

Every SVG primitive is reduced to cubic segments:

• Line → cubic with collinear controls: C₁ = P₀ + ⅓(P₃−P₀), C₂ = P₀ + ⅔(P₃−P₀).
• Quadratic → exact degree elevation: C₁ = Q₀ + ⅔(Q₁−Q₀), C₂ = Q₂ + ⅔(Q₁−Q₂).
• Circle → 4 cubics with offset k = r·(4/3)·tan(π/8) ≈ 0.5523·r.
• Arc A → center parametrization (SVG spec, appendix F.6), sliced into arcs ≤ 90°, each slice to a cubic with α = (4/3)·tan(Δθ/4).
• rect/polyline/polygon → lines (rounded rects: lines + quarter ellipses).

The d parser handles absolute/relative commands, H/V/S/T shorthands, packed arc flags and scientific notation. Result: icon = list of subpaths; subpath = chain of cubics + closed/open flag.

Instead of fighting Bézier structures of different cardinality, every subpath is sampled at N points equidistant by arc length (N = 64). Every subpath becomes a Float64Array(2N) and the structural problem disappears: correspondence is index to index.

The arc length of a cubic has no closed form; |B′(t)| is integrated with 8-point Gauss-Legendre quadrature per segment.

The refinement that drives quality: corner detection (tangent discontinuity above an angular threshold) with corners anchored as exact sample points. Remaining points are apportioned by arc length between corners — integer apportionment by largest remainder, minimum 1 interval per run, summing exactly to N−1. Effects:

• At rest the shape is exact (a check's vertex is a sample, not an approximation).
• In transit, source corners flatten smoothly and target corners sharpen — the desired behavior. Without anchoring, a morphing check visibly rounds its corners.

Closed paths anchor only corners, not the arbitrary M start point — sampling is intrinsic to the shape, so two congruent loops with different start points produce the same sample set (modulo index rotation, resolved by the circular correspondence below).

Three subproblems:

• Orientation. Both traversal directions of B are tried and scored (see the tie-break below). A hamburger line folding to the correct diagonal depends on this.
• Start point (closed paths). A closed contour has a circular degree of freedom: which index is "first". The N circular offsets × 2 directions are tried and the best post-alignment score wins. O(N²) brute force with N = 64 is ~4K operations per pair — negligible.
• Subpath matching. With p subpaths in A and q in B, pairing cost is dist(centroids) + 0.35·|ΔL| (arc lengths). Equal counts use the minimum-cost permutation (exhaustive with pruning up to 8 subpaths, greedy above); unequal counts use a surjective assignment from the large side to the small one, so no subpath ever appears or disappears out of nowhere. Leftovers are not collapsed to a point: the nearest subpath is duplicated in place and the copies separate in flight — it reads as cell division, far more natural than growing from nothing. At rest, overlapping duplicates with the same solid stroke are invisible, and the canonical snap on settle resets the count.

Before interpolating, the optimal similarity (rotation θ, scale σ, translation) mapping A onto B is computed by minimizing Σ|σ·R(θ)·(aᵢ−c_A) − (bᵢ−c_B)|². In 2D it has a closed form, no SVD:

S_xx = Σ aₓbₓ    S_xy = Σ aₓbᵧ    (centered clouds)
S_yx = Σ aᵧbₓ    S_yy = Σ aᵧbᵧ

θ* = atan2(S_xy − S_yx,  S_xx + S_yy)
σ* = [cos θ*·(S_xx+S_yy) + sin θ*·(S_xy−S_yx)] / Σ‖aᵢ‖²

The normalized RMS residual after alignment (res = √(Σ|σRa−b|² / Σ‖b‖²)) is the shape metric: ≈ 0 means A and B are the same shape rotated/scaled. This turns rotation groups into an emergent property: arrow-right → arrow-down gives residual ~0 with θ = 90° and the system chooses pure rotation on its own.

Procrustes runs per subpath (each piece rotates around its own centroid, which translates independently) — this gives local, organic motion: a hamburger folds instead of spinning as a block. A global hybrid pass then runs Procrustes over the concatenated clouds: if the global residual is ≈ 0 (< 5e-3) the whole icon is congruent and every subpath shares the same (θ, σ), so e.g. both subpaths of an arrow spin the same way.

For shapes symmetric under inversion (a straight line), both traversal orientations give residual 0 — indistinguishable to Procrustes — but produce different rotations. Naive tie-breaking can pick θ = 135° when the inverted orientation gives θ = −45° with the same end result. Each orientation is scored with:

score = res + λ·|θ|/π      (λ = 0.05)

Deformation is minimized first and, at comparable residuals, the shortest rotation wins. λ is small enough that a notably worse shape never wins just by rotating less. This is the kind of detail the math doesn't see but the eye does.

Raw coordinate lerp collapses rotations: points take the chord, the shape shrinks and shears in transit. The fix: decompose the motion into similarity + residual and interpolate each part in its natural space.

Let aᶜᵢ = aᵢ − c_A and b̃ᵢ = R(−θ*)·(bᵢ − c_B)/σ* (B brought into A's frame). Then:

P(t) = c(t) + σ*ᵗ · R(t·θ*) · [(1−t)·aᶜᵢ + t·b̃ᵢ]

  c(t)  = lerp(c_A, c_B, t)        linear translation (block transport under the global hybrid — below)
  t·θ*  = linear angle             (θ* is already in (−π, π], short path)
  σ*ᵗ   = exp(t·ln σ*)             log-linear scale (geodesic in ℝ⁺)

Exact at both endpoints (max error ~3.6e-15 in tests). Properties:

• If B is A rotated (residual 0), the bracket is constant and the motion is pure rotation.
• If there is no rotation, it's a clean coordinate morph.
• Mixed cases do both simultaneously and continuously.
• With spring overshoot (t > 1) the formula extrapolates naturally: rotation and scale overshoot slightly and come back — free juice.

Block transport (global hybrid). When the whole icon is congruent (§4), lerping each subpath centroid reintroduces the chord disease one level up: every part spins with the shared θ*, but an off-center part travels the chord — inside the arc — and drifts toward the center mid-flight. An arrow's head sagged ~1 px toward its shaft at t = 0.5 (r·(1 − cos(θ/2)) with r = 3.5, θ = 90°) while both parts rotated perfectly. Under the hybrid, each centroid rides the shared similarity around the global centroid g_A instead:

c_k(t) = c_{A,k} + t·d_k + (σ*ᵗ·R(t·θ*) − I)·(c_{A,k} − g_A)

d_k is fixed once so that c_k(1) = c_{B,k} exactly (it absorbs the tiny global residual). Every point of every subpath then moves under ONE similarity per frame: congruent icons are rigid through the whole flight, not only at the endpoints. With θ* = 0, or centroids sitting on g_A, or no hybrid at all, the formula degrades continuously to the plain lerp — no thresholds, no special cases.

In flight, each subpath is emitted as M x y L x y … (+ Z if the pair is closed↔closed; closed→open flies open, the loop opening at the cut chosen by the circular correspondence) with 2 decimals — invisible at 24px, and the frame's only allocation. On settle, the driver snaps to the canonical d of the target icon: exact fidelity at rest (real curves, not polylines), subpath count reset after duplications, and the DOM ends up identical to a static icon's. The jump is < 0.02px — imperceptible. The canonical d itself is emitted with 4 decimals, and that is a correctness choice, not a size one: arc→cubic conversion goes through trig whose last ulp differs across JS engines, and SSR hydration compares the server's bytes against the browser's — quantized emission keeps them identical, where full precision leaked each engine's ulp into the markup (a Next hydration mismatch on icons with non-cardinal arcs).

Damped harmonic oscillator over progress x: 0 → 1:

ẍ = k·(1 − x) − c·ẋ

Integrated with semi-implicit Euler at h = 1/240 s substeps (stable up to ω·h ≈ 2; with k = 420, ω ≈ 20.5 — ample margin). ~25 lines, no animation library needed.

Interruptions: when a morphTo arrives mid-flight, the plan is rebuilt from the current intermediate shape (the rendered buffers are already N points per subpath — they serve directly as source), x resets to 0 and velocity is preserved (clamped to ±14). Perceived motion is continuous; tap spam feels alive, never jumps. The spring settles and the rAF stops when |1−x| < 0.001 ∧ |v| < 0.02.

• plan(A, B) complete (resample + match + Procrustes × orientations): sub-millisecond with N = 64 — measured 0.01 ms (open paths) to 0.06 ms (closed, with circular search); 0.42 ms for a pathological 12-loops case.
• Zero numeric allocation per frame: interpolators write into the plan's preallocated Float64Arrays. Only the d string is allocated.
• Caches by reference: normalize(icon) in a WeakMap keyed by the IconNode (Lucide imports are stable references → no string keys, GC-friendly). plan(A, B) in a nested WeakMap; plans from interruptions are never cached.
• One global rAF (singleton scheduler): active instances are iterated, settled ones unregister. A hundred icons morphing at once = one loop.

CI budget (size-limit, gzip) as an anti-regression tripwire — gates carry ~10% headroom over what's measured; if a real capability needs more, the number is renegotiated, but growing unnoticed is not accepted:

The adapters entry carries two kinds of gate: the aggregate (allowed to grow as adapters are added) and one per named export, which pins each adapter as independently tree-shakeable — if two adapters ever couple, the per-export gate trips in CI. Standalone, { svgToIcon } measures 3.34 KB because auto-fit drags the core's lowering machinery; an app that already imports the core shares that chunk and pays ~0.4 KB incremental.

bun run play   # → http://localhost:3000

38 real icons from Lucide (IconNode imported from the package), Feather and Tabler; per-pair readout of the detected similarity (θ, σ, residual), polar/linear modes for comparison, spring presets and a t scrubber. Feather/Tabler data is extracted from the npm packages with bun playground/extract-vendor-icons.mjs.

bun test          # 156 tests / ~13,700 asserts
bun run typecheck # strict ×6: core+dom without lib DOM, playground, react, react-native, vue, svelte
bun run format    # biome
bun run build     # tsdown → dist/ + scripts/build-svelte.ts (svelte entry)
bun run size      # size gates

Icons shown in the playground belong to their authors: Lucide (ISC), Feather (MIT), Tabler (MIT).

MIT © Guillermo — made by guillermolg.com