A drop-in lipsync renderer for stylized character profiles — no rigged avatar, no ML inference, no server roundtrip. wireface analyses voice or TTS audio in the browser and drives a wireframe face mesh through 30 expression channels — 15 visemes from the Oculus Lipsync phoneme alphabet plus 15 ARKit-style face blendshapes (jaw, lips, blinks, eye gaze, squint, brows, nose, head rotation). Each character is one JSON preset — mesh resolution, colors, depth fade, glow, mood weights, channel gains — saved from the bundled editor and replayed by a single JS file. One canvas, one peer dependency (Babylon.js v9), one createWireface(canvas) call. Each instance is fully independent, so a page can host as many talking heads as it has audio sources.
- repo: https://github.com/styk-tv/wireface
- npm: https://www.npmjs.com/package/wireface
- one source file:
wireface.js - two runnable examples:
examples/demo.html(consumer),examples/editor.html(preset authoring) - two preset samples:
examples/presets/ - preset format: plain JSON (drop a
.jsonsaved by the editor onto a panel) - audio: any browser-decodable file (
mp3,wav,ogg,m4a,flac,aac,opus)
npm install wirefaceTwo flavors are published — pick by file extension:
| Flavor | jsDelivr | unpkg |
|---|---|---|
| indented (readable, debuggable) | https://cdn.jsdelivr.net/npm/wireface@1/wireface.js | https://unpkg.com/wireface@1/wireface.js |
| minified (smaller, production) | https://cdn.jsdelivr.net/npm/wireface@1/dist/wireface.min.js | https://unpkg.com/wireface@1/dist/wireface.min.js |
Pin to a major (@1), a minor (@1.0), or an exact version (@1.0.0) — latest works too but isn't recommended for production.
wireface has one peer dependency: Babylon.js (babylonjs >= 6.0.0). Load it from CDN or bundle it yourself before wireface.js runs.
<!-- Babylon.js v9 as a global, loaded BEFORE wireface.js --><scriptsrc="https://cdn.jsdelivr.net/npm/babylonjs@9/babylon.js"></script><!-- pick one — minified for production, indented for debugging --><scriptsrc="https://cdn.jsdelivr.net/npm/wireface@1/dist/wireface.min.js"></script><canvasid="face" style="width:480px;height:480px"></canvas><script>constwf=createWireface(document.getElementById('face'));// Optional: load a saved preset (see "Preset JSON" below)fetch('preset.json').then(r=>r.json()).then(p=>wf.loadPreset(p));// Pull an audio file (or use a File from a drop / file input)fetch('voice.mp3').then(r=>r.blob()).then(blob=>wf.loadAudio(blob));// Drive itwf.setMood('happy');wf.setLoop(true);wf.play();</script>That's the entire surface. Everything else is presets, channels, and moods.
createWireface(canvas: HTMLCanvasElement,options?: object)→WireFaceinterfaceWireFace{// audio + presetloadAudio(file_or_blob): Promise<void>loadPreset(presetObject): void// the JSON saved by the editorsetRenderConfig(partial): void// apply one-or-more renderConfig fields livegetRenderConfig(): object// snapshot of current renderConfig// transportplay(fromOffset?: number): voidpause(): voidstop(): voidsetLoop(bool): voidisPlaying(): booleangetDuration(): number// seconds (0 if no audio loaded)getPosition(): number// seconds since playback start// expressionsetMood(name): void// 'neutral'|'happy'|'sad'|'angry'|'fear'|'surprise'|'sleep'getActiveMood(): stringsetChannel(name,value): void// manual override (auto-releases after ~1.2s)setChannelGain(name,gain): voidgetChannel(name): number// current smoothed value (0..1 or -1..1)getChannelTarget(name): number// latest target before smoothinggetChannelGain(name): numbergetChannelNames(): string[]getMoodNames(): string[]// camerasetView(view): void// 'front'|'three-q'|'profile'|'orbit'// teardowndispose(): void}Multiple instances are independent — see examples/demo.html for a two-panel example with per-panel drag/drop, transport buttons, and mood pills.
A preset is one plain object — exactly the shape produced by the editor's save button. Two examples ship with this repo:
examples/presets/asset-th5rha.json— dense21×21red-wire face, glowy iris, deep eye socketsexamples/presets/asset-ydr7de.json— sparse7×9low-poly blue-wire face, exaggerated mouth scale
| Field | Type | Meaning |
|---|---|---|
version | string | preset format version (e.g. "wireface-v010") |
createdAt | ISO string | when the preset was saved |
renderConfig | object | mesh + render + style — see below |
channelGains | object | per-channel gain (0..N); scales how much each input drives the face |
channelGainSliders | object | UI-side slider position; loadPreset reads channelGains |
activeMood | string | one of neutral, happy, sad, angry, fear, surprise, sleep |
moodTargetWeights | object | per-mood target weight 0..1 (sum is normally 1) |
audioName | string | hint of the audio that was paired with this preset (informational) |
lineColor / pupilColor | hex | top-level color overrides applied after renderConfig |
id / name | string | UI identifier |
| Group | Field | Range | Notes |
|---|---|---|---|
| mesh | meshCols / meshRows | int | grid resolution; rebuild on change. e.g. 21×21 (smooth) or 7×9 (low-poly) |
mode | "wire" / "solid" / etc. | render mode | |
meshVisible | bool | show the underlying grid | |
minimal | bool | thin line overlays for mouth/eyes/brows/nose | |
fragment | bool | break the wire mesh into per-tri fragments | |
flipFace | bool | mirror the face left↔right | |
lineThickness | 0.1..2 | overlay tube radius | |
| face shape | scaleNose / scaleEyes / scaleMouth | 0..2+ | per-feature scaling |
spread | 0..1 | how widely shapes spread across the mesh | |
reactivity | 0..1 | overall morph amplitude in response to audio | |
lipVertAmp | 0..N | vertical amplitude of lip motion | |
lipPressForce | 0..N | "press" force for closure visemes (PP, MM-like) | |
jawExtend / mouthGrow | offsets | static mouth shape biases | |
| holes / depth | mouthHole / eyeHoles | bool | cut holes in the mesh |
mouthHoleSize | ~0.4..2 | mouth-hole ellipse radius multiplier | |
eyeHoleSize | ~0.4..2 | eye-hole ellipse radius multiplier (per-eye) | |
eyeDepth | 0..5+ | how deep eye sockets recess into the parametric warp | |
mouthAnchor / eyeAnchor | bool | pin anchor rings to track shape changes | |
overlayTracksMesh | bool | minimal-line overlays follow mesh deformation | |
| style | lineColor | hex | wire color |
baseColor | hex | mesh base color | |
fadeColor | hex | depth-fade target | |
depthFade | 0..1 | depth attenuation strength | |
pupilColor / irisColor / irisSize | hex / 1..4 | eye styling | |
browColor | hex | brow line color | |
glow | bool | bloom-style glow on emissive lines | |
pupils | bool | render pupils at all | |
| mood | moodTransitionTime | seconds | crossfade time between moods |
Channels are the live signal surface. Audio analysis writes them; setChannel() overrides them; channelGains scales them; the renderer reads them every frame.
viseme_sil viseme_PP viseme_FF viseme_TH viseme_DD viseme_kk
viseme_CH viseme_SS viseme_nn viseme_RR viseme_aa viseme_E
viseme_I viseme_O viseme_U
jawOpen mouthSmile mouthPucker
eyeBlinkLeft eyeBlinkRight eyeLookH eyeLookV eyeSquint
browInnerUp browOuterUp browDown
noseSneer
headRotateX headRotateY headRotateZ
Manual overrides via setChannel(name, value) lock the channel for ~1.2s, then auto-release back to audio-driven control.
Moods are convenience presets that fade across a fixed subset of channels (browInnerUp, browOuterUp, browDown, mouthSmile, eyeSquint, noseSneer, eyeBlinkLeft, eyeBlinkRight):
| Mood | Effect |
|---|---|
neutral | all-zero |
happy | brows up, smile, slight squint |
sad | inner-brow up, slight brow down, frown, half-blink |
angry | brow down, frown, sneer, squint |
fear | brows up, frown, eyes wide |
surprise | brows up, jaw open |
sleep | eyes closed, brows down |
Crossfade time is controlled by renderConfig.moodTransitionTime.
setView(view) snaps the camera; 'orbit' engages a slow auto-orbit.
| View | Camera |
|---|---|
front | dead-on |
three-q | three-quarter |
profile | side |
orbit | slow continuous orbit |
| File | Purpose |
|---|---|
examples/demo.html | minimal library consumer — two side-by-side instances, drag-drop preset/audio, transport buttons, mood pills. Use this as your reference for embedding wireface into your own page. |
examples/editor.html | full preset authoring UI — sliders, knobs, color pickers, mood weights, save/load. Currently self-contained (ships its own copy of the rendering pipeline) for offline authoring; produces the .json shape the library consumes. |
To run them locally:
npm run demo
# → opens a static server at http://localhost:5173 — visit /examples/demo.html# or /examples/editor.htmlThen drop the bundled presets onto each panel of the demo:
examples/presets/asset-th5rha.json → left
examples/presets/asset-ydr7de.json → right
…and drop any .mp3 of speech on top of either panel.
loadAudio()decodes viaAudioContextand wires through anAnalyserNode.- Each frame, the renderer reads frequency-band energies and translates them into viseme channel targets (
viseme_aa,viseme_O, …) plus jaw open / lip press / smile. - Active mood blends in across mood-channels with
moodTransitionTimecrossfade. - Idle micro-motion adds breathing, blinks, and head sway when no audio is playing.
- Per-channel smoothing prevents the face from snapping (visemes get heavier smoothing, brows lighter — see
SMOOTHtable inwireface.js). - Final channel state drives mesh morphs + minimal overlay line meshes (mouth lips, eyelids, brows, nose).
Issues, PRs, and ideas are very welcome. Anything that's fair game:
- new viseme / expression channels, or better channel mapping for non-English phonetics
- mood blends, idle micro-motion presets, eye-look behaviors
- performance / memory tweaks (smaller draws, tighter geometry rebuilds, WebGPU path)
- alternative renderers (svg, 2d-canvas) sharing the same channel surface
- editor UX: keyboard shortcuts, undo, multi-channel keyframing
- docs, typos, clearer onboarding
Open an issue or send a PR — small ones are perfect.
The single most useful contribution is a preset of your own face. Take
the editor, tune sliders + colours + channel gains
until you've got something that feels like a character to you, then hit
+ save preset → ↓ download and submit a PR adding the JSON to
examples/presets/ with a one-line README entry
describing the vibe (e.g. "scared crocodile", "neon priest", "low-poly
robot"). Author yourself in the filename or a comment — we'll keep
attribution intact.
The two presets bundled today (asset-th5rha, asset-ydr7de) are
starting points, not the canon. The library's whole point is character
variety, and the only way that scales is community-built preset packs.
Same kind of generosity that built the demoscene.
Built something with wireface? Please share it. There's a dedicated space:
- 👉 GitHub Discussions → Show & Tell — drop a screenshot, a video, a CodePen, a deployed URL, or a
.jsonpreset you're proud of. Other people's presets are the best kind of docs.
If you build a public tool / site / experiment with it, a star on the repo is a small thank-you that makes a real difference.
The channel grammar is the union of two well-established blendshape inventories. Pick one to read end-to-end and the other will look familiar.
The 15 viseme channels (viseme_sil, viseme_PP, viseme_FF, viseme_TH,
viseme_DD, viseme_kk, viseme_CH, viseme_SS, viseme_nn, viseme_RR,
viseme_aa, viseme_E, viseme_I, viseme_O, viseme_U) come from the
Oculus Lipsync SDK's phoneme alphabet. The same set ships with HeadTTS,
Kokoro, ovrlipsync, and most "phoneme alignment" pipelines. Audio FFT
estimates the active viseme each frame and the renderer weighted-sums the
(open, width, round, smile, lipPress, jaw) targets.
- Oculus Audio SDK — Viseme Reference — the canonical 15-viseme table plus example audio cues per phoneme.
- Oculus Lipsync overview — what the SDK does and why the 15-viseme split.
The remaining 15 channels (jawOpen, mouthSmile, mouthPucker,
eyeBlinkLeft/Right, eyeLookH/V, eyeSquint, browInnerUp/OuterUp/Down,
noseSneer, headRotateX/Y/Z) are a curated subset of Apple ARKit's full
52-blendshape inventory, named to match. ARKit blendshapes are the
de-facto industry standard for face animation since iPhone X / TrueDepth;
glTF face blendshapes, MetaHuman, NVIDIA Audio2Face, and Ready Player Me
all use compatible naming.
- ARKit
BlendShapeLocationreference — full 52-blendshape list with semantic names + face-region grouping. - ARKit Face Tracking overview — what ARKit's blendshapes mean and how they're estimated from the TrueDepth depth+IR camera.
A renderer that only knew Oculus visemes would lipsync audio but couldn't
emote. A renderer that only knew ARKit blendshapes would emote but
lipsync poorly (ARKit's mouth blendshapes are continuous 0..1 dilations,
not phoneme-aligned). Combining both gives mouth that lipsyncs from
audio + face that emotes from explicit channels — they layer additively
in computeMouthParams() so both signals can drive the face simultaneously
without conflict.
The full grammar (channel names, ranges, smoothing, drive maps, preset
JSON shape, invariants) is in
SPEC.WIREFACE.GRAMMAR.v1.0.md.
Babylon.js setup details (multi-instance scenes, multi-canvas pages,
playground integration) are in
SPEC.WIREFACE.BABYLON.9.0.md.
MIT © 2026 Peter Styk — see LICENSE.