Parameters, with provenance¶

Verbatim output of the liquid-glass-optical-spec research run. Preserved unedited; see README for what it is and how far to trust it.

All lengths are authored in dp and converted to px at the host (dp × density). Provenance tags: [apple] stated by Apple; [measured] third-party side-by-side measurement against a real iOS 26 device; [derived] follows from optics/geometry; [ours] a taste constant with no external source. Apple publishes zero numeric rendering parameters, so nothing below is tagged [apple] on a number — only on a direction or a rule.

Optics¶

Uniform	Value	Unit	Provenance
`uIor`	1.5	index	[derived] window glass; also the default in sorrell.info, kube.io, ybouane, Compose-Internals. Because `bend` is normalised by its own rim value, IOR sets the shape of the falloff, not its magnitude — raising it to 2.0 changes the curve by <4 %.
`uBevelPower` (p)	2.0 default, 4.0 tight	exponent	[derived]. p = 2 is the true circular arc; p = 4 is the squircle-matched bevel (Codrops `bevelPower`, kube.io `(1−(1−x)⁴)^(1/4)`). Measured normalised falloff at n = 1.5: p = 2 → 1.000 / 0.645 / 0.522 / 0.386 / 0.302 / 0.188 / 0.086 / 0.001 at e = 0 / .05 / .1 / .2 / .3 / .5 / .75 / 1. p = 4 → 1.000 / 0.600 / 0.416 / 0.226 / 0.130 / 0.041 / 0.005 / 0.000. I default to p = 2 because it lands between the two independently tuned shipped falloffs (Kyant0's `1−√(1−x²)`, RN's `rim²` = 0.25 at mid-band, this repo's `t³` = 0.125) and is the physically derived case; p = 4 concentrates ~95 % of the displacement into the outer third of the band and needs a proportionally wider `uBand`.
rim shift bound	√(n²−1) = 1.118 at n = 1.5	ratio	[derived]. The exact Snell form self-bounds here, so no slope clamp is required. Numerically the profile reaches 0.9885 (p = 2) / 1.0601 (p = 4) at the e = 0.002 clamp; normalisation divides it out.
Schlick `F0`	0.04 at n = 1.5 (0.0200 at 1.33, 0.0338 at 1.45, 0.0426 at 1.52)	reflectance	[derived] `((n₁−n₂)/(n₁+n₂))²`. Known error vs exact Fresnel: −0.019 absolute (≈−21 % relative) at θ = 60–65°, +0.036 at 85°. Accepted; the term is a stand-in for an environment we do not have.
gradient epsilon	`clamp(uBand·0.3, 1, 16)`	px	[derived + graphics-writeup]. 1 px produces a one-pixel medial-axis seam down every pill.
`axisFade`	`smoothstep(0.15, 0.80, \|g\|/(2·eps))`	—	[graphics-writeup]
C1 seam fade	`smoothstep(1.0, 0.75, e)`	—	[ours], inner quarter of the band
rim guard	`smoothstep(0, 1.5, depth)`	px	[ours]. 1.5 px, below perception; only stops the outermost antialiased pixel fetching from beyond coverage. Do not widen this to 12 px — at a 14 dp band that is a visible dead ring.

Displacement magnitude — the largest disagreement in the corpus¶

Published tuned values for peak rim displacement, all on ~48 dp controls: Kyant0 24 dp (refractionAmount, refractionHeight 12 dp); react-native-liquid-glassmorphism 9 dp regular / 11 dp clear, ×1.35 with the refraction flag, explicitly eye-matched against iOS 26; this repo today 22 dp regular / 9 dp chrome. That is a 2.7× spread.

I chose the 9–18 dp family, i.e. the device-matched end, because (a) it is the only figure any source claims to have compared side-by-side against an iOS 26 device, and (b) it has a falsifiable failure mode at the other end: over 17 sp body text, rim displacement above ~16 dp yields glyphs that are still legible but smeared along the rim, which reads as a rendering defect rather than as glass — the same finding that forced this repo's Chrome preset to 9 dp.

Presets (dp unless noted)¶

	Regular	Clear	Chrome	Thick
`uBand`	26	28	14	22
`uDepth`	14	16	9	12
`uBevel`	4	4	3	5
band : bevel	6.5 : 1	7 : 1	4.7 : 1	4.4 : 1
`uBlur` (in-shader)	10	4	0	16
host pre-blur	0	0	14	0
`uDispersion`	0.018	0.030	0.010	0.015
`uMirror`	0.18	0.22	0.08	0.14
`uTint` α	0.10	0.04	0.11	0.13
`uAdaptive`	0.65	0	0.90	0.80
`uLegibility`	0.6	0	0.9	0.5
`uFlip` allowed	yes if small	no	yes if small	no
`uSpecular`	0.55	0.70	0.50	0.45
`uSpecularPow`	6	6	6	7
`uFresnel`	0.10	0.14	0.08	0.08
`uInnerShadow`	0.07	0.05	0.06	0.10
dimming layer (host)	0	0.35	0	0

Band : bevel ratio 4–7 : 1 — [derived]; the refraction band is an independent artistic length, not the lit bevel.
uDispersion 0.018 — [measured]; 0.06–0.09 (this repo today) is ~4× the device-matched value and starts reading as channel breakup on a high-contrast edge.
Clear's zeros are [apple]: "Clear… does not have adaptive behaviours. It is permanently more transparent." Clear therefore does not tone-map, does not flip, does not auto-tint content, and requires the dimming layer.
Dimming 0.35 — [apple], the single numeric parameter in the entire HIG ("a dark dimming layer of 35 % opacity" when the underlying content is bright). The SwiftUI Glass.clear code sample uses .black.opacity(0.3); I take 0.35 as the specified value and treat 0.30 as sample drift. Skip it entirely when content is already dark or when the player supplies its own. For a small footprint, dim the element's footprint plus a falloff rather than the whole content layer [apple].
Tint alphas — [ours], bracketed by [measured] 0.18 regular / 0.047 clear from the RN library (whose pipeline also carries a separate adaptive frost term, so the numbers are not directly interchangeable).

Adaptation, size and the flip¶

Quantity	Value	Provenance
element size factor	`clamp((min(w,h)_dp − 56) / (320 − 56), 0, 1)`	[ours]. 56 dp control = 0 (small, clear), 320 dp surface = 1 (large, opaque).
size → lens gain	`mix(0.85, 1.25, scale)`	[ours], direction [apple] ("more pronounced lensing and refraction" as glass morphs larger)
size → tint gain	`mix(0.9, 1.25, scale)`	[ours], direction [apple] ("larger size is more opaque, smaller is clearer")
size → inner shadow gain	`mix(0.9, 1.15, scale)`	[ours], direction [apple] ("deeper, richer shadows")
flip permitted when	size factor < 0.5 (min edge < 188 dp)	[ours]; the rule that small flips and large does not is [apple], the threshold is not published anywhere
flip decision	host: mean Rec.709 luma of a 6×5 sparse read of the captured backdrop under the element, ≤10 Hz	[ours]; the 30-sample grid is [measured] practice
flip hysteresis	enter dark-on-light at L > 0.60, leave at L < 0.44	[ours]. Apple publishes no threshold, no sampling region, no cadence and no hysteresis band; a search of HIG Materials, HIG Color, Adopting/Applying Liquid Glass, UIGlassEffect, SwiftUI Glass and WWDC25 219/356 returns zero numeric thresholds.
flip crossfade	180 ms linear on `uFlip`	[ours]
tone map excursion	value `mix(1.10, 0.88, L)`, saturation `mix(0.88, 1.06, L)`, both lerped by `uAdaptive`	[ours], direction [apple]
legibility lift	`+0.25 · uLegibility · contrast`, `contrast = clamp(\|luma(sharp)−luma(soft)\|·4, 0, 1)`	[ours], direction [apple]
strength ceiling	0.95	[ours] — the material must never become opaque

Lighting¶

Quantity	Value	Provenance
key light	unit vector (−0.35, −0.94), y negative = above	[ours]. Above and slightly left is the direction a highlight must come from for a surface to read convex.
key lobe	`pow(max(n·L,0), 6)` over the bevel band	[ours]
counter lobe	`pow(max(−n·L,0), 9.6) × 0.35`	[ours]; two-sidedness is [apple]-consistent and matches LiquidGlassKit's doubled-bearing glare and Kyant0's `pow(\|dot\|, falloff)`
edge line	outermost 2 px, `0.07 + 0.5·max(n·L, 0)`	[ours]
inner thickness line	band `[uBevel, 2.2·uBevel]`, strength `uInnerShadow`	[ours]

Interaction¶

Quantity	Value	Provenance
press luminance lift	+0.06 at peak	[measured] against iOS 26; 0.2 gives ~50 % lift and reads as a flash
glow falloff	Gaussian, σ = min(w, h) px, from the touch point	[measured] form, [ours] σ
touch magnifier	sample offset −0.17 × (local − touch) × falloff	[measured]
glow spill to neighbours	container-level: evaluate the Gaussian in container space so neighbouring members light up too	[apple] behaviour ("spreads… onto any Liquid Glass elements nearby")

Container / merging¶

Quantity	Value	Provenance
`MAX_GLASS_MEMBERS`	8	[ours]; SkSL needs an unrollable loop, so the cap is compiled in
`uMerge` (smin k)	k = 2 × spacing_px	[derived]. With the un-normalised polynomial smin (`h = clamp(0.5+0.5(b−a)/k)`, max inflation k/4), two shapes whose edge gap is g meet at the midline when g ≤ k/2. Setting k = 2·spacing makes the silhouette fuse exactly when the nearest edges are `spacing` apart, which is Apple's stated semantics. Note the widely-copied `k := spacing` is wrong by 2× under this form and by 4× under iq's normalised form.
default `spacing`	40 dp	[apple] — the value in Apple's own `GlassEffectContainer(spacing: 40.0)` sample
container spacing vs layout spacing	container `spacing` ≤ interior stack spacing	[apple] — a larger container spacing "causes Liquid Glass effects to blend together at rest"

Host contract¶

Quantity	Value	Provenance
`uPad`	`uDepth·1.4 + max(uBlur, hostBlur)` px	[derived]. Covers the 1.25× size gain and the 1.03× dispersion reach with margin.
`uBackdrop`	intersection of (padded layer) ∩ (recorded backdrop), inset 0.5 px	[derived]; half a pixel keeps bilinear filtering off both boundaries
`uBase`	the app's opaque ground colour	[derived]
shadow alpha	0.14 over solid light ground → 0.30 over text	[ours]; the inversion (more over text, less over solid light) is [apple]
shadow blur / offset	`18 dp × (0.6 + 0.6·scale)` / `y = 4 dp × (0.5 + scale)`	[ours]
shadow update cadence	≤10 Hz, same sparse backdrop read as the flip	[ours]
scroll edge — soft	band = overlap + 12 dp, opacity ramp 0 → 0.9 quadratic, plus progressive blur	[ours]; soft/gradient is [apple]
scroll edge — hard	uniform across the full toolbar height incl. pinned accessory	[apple]
scroll edge — dark mode	switch from fade-to-background to a multiply dim of 0.18	[ours]; the switch is [apple]
Reduce Transparency	`uFrost = 1` (+0.35 strength) and `uBlur × 1.6`	[ours]; "frostier, obscures more" is [apple]
Increase Contrast	`uContrast = 1` → 88 % toward black/white + contrasting border	[ours]; "predominantly black or white… contrasting border" is [apple]
Reduce Motion	disable press scale/bounce and the touch magnifier; keep the glow at 50 %	[apple] ("decreases the intensity of some effects and disables any elastic properties")

Cost¶

23 backdrop fetches worst case per covered pixel: 19 (scatter) + 3 (dispersion) + 1 (echo). Zero extra fetches for local contrast. Falls to 4 when uBlur·e < 0.5 (i.e. across the whole rim band, and everywhere in the Chrome preset). Measured on the existing 22-fetch ancestor of this shader: Pixel 7 at 120 Hz, two glass elements over live app content, 4.36 % janky frames with the shader on vs 4.96 % with it forced off — the app's idle redraw dominates. Budget ≤3 simultaneous glass elements and one container per screen; Apple gives the same rule qualitatively ("limit the use of Liquid Glass effects onscreen at the same time").