Skip to content

Roadmap

For work completed through13 October 2026 and the remaining delivery groups, use the current visual status. The source contains native.NET calm feedback and a standard WinForms-control backdrop adapter in addition to Compose and portable paths. The historical strategy below does not supersede the current requirement checklist.

Historical strategy, September 2026. Implementation notes below describe the original audit. For current delivered behavior, use implementation status. The Atlas stabilization now includes continuous magnifiers, calm interaction, desktop visual evidence and portable WebGL/Skia paths; native bindings and full iOS equivalence remain open. Counts and competitor comparisons below are historical observations, not a fresh survey.

Written 2026-09-10, after reading the source of every liquid-glass implementation that does real optics. The audit is in research/competitive-landscape.md; this is what to do about it.

The strategic position, in one paragraph. Two libraries dominate this space and both refuse to ship components: Kyant0's says so in one sentence in its README, and Haze has declined for three years. The only project that ships components and real optics has 26 stars and is one month old. Meanwhile the only measured tab bar in the field is ours. So the defensible ground is optics fidelity plus a modifier that converts anything, and the thing that erodes it is reach: both leaders build for iOS, macOS, web and desktop, and we build for Android and desktop JVM. Fix reach or none of the rest matters.

P0 — a defect two independent measurements agree on

Done. The direction-independent floor is gone from the edge line, which is now entirely directional, and GlassStyle.edgeShadow draws the dark contour that defines a shape where nothing is lit. Measured back on device: the lens's side rim reads a dark step at 21 against an interior of 42, where the reference reads 24 against 41. edgeLight rose 1.14x on the two affected presets to hold the lit line where it already measured correctly.

Two faults surfaced doing it, both now lessons. A contour drawn on the outermost pixel is invisible over a dark ground, because coverage there is partial and the result is composited against whatever lies outside; it has to sit inside the ramp. And the device harness was reading the bar's edge rather than the lens's, because difference-based detection spans the whole bar when every icon recolours on press.

P1 — take what the competition proved

The analytic gradient: done. For a closed-form shape whose refraction band stops short of the medial axis — which is the common case and both of our own presets — the surface normal now has a closed form, saving four field evaluations per pixel. The wide central difference stays for sampled paths and for bands that do reach the axis, because that is the only case its width was ever for. Frame cost on the tab bar fell from 14ms to 13ms at p50 and 32ms to 29ms at p99, with every measured parity number unchanged.

The fingertip bulge: already present, and this item was an error. The roadmap asked for something the shader has had all along — touchFall displaces sampling toward the finger, on a Gaussian whose sigma is the panel's short edge. The only real difference from QWEA0's is that theirs is tighter and ours is deliberately broad, so a press on one member of a container lifts its neighbours. Nothing to do.

The dome field: deferred, with a measurement. Haze replaces the distance field with a dome field for elongated shapes, which is a cleaner answer to the medial-axis collapse than our confidence fade. It engages only when the refraction band reaches the inradius, and a rendered test at exactly that worst case now shows no seam with what we already have: the row on the axis sits within 12 counts of its neighbours over a hard vertical stripe pattern. Our presets are nowhere near that regime — a band of 42px against an inradius of 108 on the bar. The open question is not the seam but how much refraction strength the fade costs down the spine of a very wide pill, and that is worth measuring before adding a second field.

The tint model: deferred, deliberately. QWEA0 multiplies by the tint and adds a scattering term, which preserves the backdrop's own light and shade where a lerp flattens it. Ours is a lerp toward a tone-mapped colour — but its strength is fitted to measurement: 0.36 is the value that both lifts pure black by 20 and passes 64% of white text, the two-measurement fit that pins the material. Swapping the model would break that fit for a theoretical gain. If it is done, it belongs behind a knob defaulting to the current behaviour, the way edgeShadow was.

P2 — reach, which is the real gap

Consumable: done. The library is its own repository, with the app it was built for consuming it as dev.shayxo.liquidglass:liquidglass:0.1.0 from a Maven repository rather than as a module. Android and JVM variants, Gradle module metadata, sources jars and a POM, under Apache-2.0; the licence is the canonical text and the README carries a statement of non-affiliation. 0.1.0 has been on Maven Central since 2026-09-10, published by the tag-driven workflow (publishing.md, which also records what the first release taught). The coordinates are now fixed: a group or artifact id cannot change without orphaning consumers.

Targets: blocked on hardware, not on design. Kotlin/Native cannot build iOS or macOS targets from Windows, so adding them here would mean shipping code nobody has compiled. The desktop path already runs the same shader through Skia, so the work when a Mac is available is mostly source set plumbing: rename jvmMain to a shared Skia source set and let the native and web targets depend on it. Two things to know before starting. Compose Multiplatform 1.11 turned the non-Android shader into a Compose wrapper type rather than a raw Skia one, and our desktop sources import Skia directly today. And the web targets can be built from here, so they are the sensible first move.

The platform backdrop: not started. Android 17 QPR2 makes public a render node method that applies an effect to the pixels already drawn behind a node, which is a correct backdrop at no capture cost. It does no optics. Haze wires it in behind a feature flag with a capture fallback; we should do the same. Not testable on this device, which is on Android 16.

P3 — the moat is the modifier, and components are its proof

A slider and a toggle were built here, and both were deleted. They are the reason this section is now titled the way it is. The direction that removed them: the point of the library was never a catalogue of glass widgets, it is that one modifier on any composable (Modifier.glassEffect() since 18 October, Modifier.liquidGlass() before) turns it into the material — a stock Material 3 Card with a transparent container and one modifier — and then every component comes for free. A catalogue competes on breadth, which is a race, and it dates the moment a host wants a control the catalogue does not have.

What replaced them was the ambient state — LiquidGlassScene, ProvideLiquidGlassState, and overloads of liquidGlass() and liquidGlassSource() that take no state (all since removed) — and, since 18 October, LiquidGlassTheme at the root with Modifier.glassEffect, which needs no state at all; MaterialConversionActivity in the sample is the standing proof. So the API reference has no slider in it, and should not.

A component earns its place here only when the modifier alone cannot express it, which means it carries behaviour rather than appearance. Two do:

  • GlassTabBar, because a selection lens that stands proud of its bar, looks through it, is dragged 1:1 and commits on release is behaviour, not a style. Every default in it is a measurement.
  • GlassEffectContainer (LiquidGlassContainer before 18 October), because fusing panels is a change to the distance field shared by all of them, which no per-element modifier can reach.

By that test the open candidates are a bottom sheet, a navigation bar and a search field, each of which has motion the modifier cannot supply on its own. A toggle and a floating action button do not qualify: they are a shape and a modifier. And for the tab bar, the behaviour Apple ships and we do not — minimize on scroll, a bottom accessory that moves inline when minimized, and a dedicated trailing search tab.

P4 — below Android 13

Everything good in this field needs API 33. We degrade to a tinted surface with rim lighting, which is honest but plain. QWEA0 does real refraction and chromatic aberration down to API 24 through a NEON path in C++. That is a large piece of work and it is the difference between a library that looks right for most users and one that looks right for all of them. Schedule it after P2, not before.

The measurement programme, revised

Parity to a single number is no longer meaningful. The 2026 release replaces the earlier two-position toggle with a continuous transparency slider whose default is its midpoint, so the material's opacity is now a user setting with a range. What stays measurable, and therefore worth measuring, is everything the slider does not move: the geometry of the rim, the direction it samples, the lighting lobes, the timing of formation and subsidence, and the shape of the resting state.

Three specific captures would settle open questions:

  1. The same tab bar at both ends of the slider, to separate what the material does from what the setting does. This resolves the discrepancy our measurements doc records but cannot explain, where two captures of the same bar differ by more than a factor of two in body brightness.
  2. A glass element overhanging a hard-edged boundary, photographed on a device known to be on the 2026 release. That is the experiment that discriminates sampling direction, and it is the claim on which we differ from every other implementation.
  3. The 2026 edge treatment. Done on 19 October: the Ios27 preset's guessed numbers were replaced with iOS 27's measured rim (the outer contour, the lit edge's peak and half-depth, no dispersion).

Naming and publishing

Apple has not registered "Liquid Glass" as a trademark; it is absent from Apple's own published list, updated two weeks before this was written, fifteen months after launch. No takedown has ever been filed over a reimplementation, and the largest projects in the field have carried the name unchallenged for over a year, including one maintained by a consultancy with legal review.

The one real constraint is assets. Apple's SDK agreement licenses its system images only for building apps for Apple platforms, which rules out shipping SF Symbols, extracted textures or lifted shaders. Our shader is ours, so we are clear.

Follow the pattern the market leader already set: a descriptive repository name, a distinctive artifact name, the Apple term used only descriptively in prose, and a short statement of non-affiliation. None of this is legal advice, and the naming question is worth thirty minutes of a trademark attorney's time before the library carries commercial weight.

After the calibration study (0.2)

Open, in order of value: the wide term's asymmetry (the phone lifts more over a brightening neighbourhood than it drops over a darkening one; a luminance-adaptive lift is the candidate); the sheet material's own dependence on Tint Amount (confounded in the captures by a notification dim); a toolbar over a flat region of the target, to split the toolbar role's opacity from its fine share (its stripes fix only the product); elements under 15 dp of corner radius; widget glass; and whether the highlights respond to device tilt. Each needs a capture before it needs code. Added 2026-09-12: whether the fine term's sharp axis is the screen's x or the element's long axis (every element it was measured on is a horizontal bar; a vertical one over stripes settles it); the pill's one-pixel dark contour; the light appearance of the tab bar's indicator and lens; and, on the performance side, a system trace of the UI thread, since caching the render effects across frames moved nothing (six panels 6.2 ms before and after, Vitals 2.9 ms), so the per-panel cost is in the layer records or in layout.