haze 0.5.0

Progressive (gradient) blur for Flutter: a true variable-radius Gaussian that fades from any edge, with an optional matching tint wash.

0.5.0

• Breaking: the API is Haze(edge, sigma, tint, enabled, child). Removed tintOpacity (use the tint's alpha), tintAdaptivity, plateau, blurPlateau, falloff, blurCurve, borderRadius, debugProfile, Haze.resolve, Haze.fadeRoom, Haze.falloffAt and the ceiling helpers. The hold lengths (41% blur, 35% tint) are fixed at the values measured off the system effect, and sigma is fitted to the height automatically (Haze.fitSigma).
• Geometric sigma ramp. Replaces the squared profile, which parked sigma high over most of the fade and spent the whole visible change in a short stretch: it read as a jump to full blur, and large sigmas looked alike.
• Tapered kernel. The Gaussian is cut at 3.5 sigma, not rounded, and lands on exactly zero there. ceil(3 sigma) switched a ~1% tap on at every row where it crossed an integer, stacking faint lines down the fade. The tap cap is raised from 64 to 128, and taps are never skipped (the old stride shimmered under scrolling content).
• Premultiplied-correct tint, and dither on the final colour instead of only the tint's alpha.
• Fallback bands meet edge to edge on physical pixels (the 0.5pt overlap drew a line at every boundary) and follow the shader's curve.
• No second, unqualified shaders/haze.frag lookup: another copy with a different uniform layout would load silently.

0.4.4

• debugProfile: true paints the two profiles instead of the effect — red the sigma's, green the scrim's — so a curve problem can be told apart from a coordinate one by looking. Both must reach black exactly at the rectangle's far side.

0.4.3

• The line at the start of the fade is gone. falloff warped the position along the ramp with pow(d, 1 / falloff), which has an infinite derivative at the junction, and the smootherstep could not absorb it: the composed profile is about d^(3/falloff) there, whose second derivative vanishes only below 1.5, sits at a non-zero constant AT 1.5 — the default — and diverges above it. A curvature break on a slow ramp is a Mach band, so every band drawn at the default settings had a faint line across it exactly where the plateau gave way, which read as the hard start of a plain BackdropFilter. The warp is a rational bias now: same endpoints, same shape between them, finite non-zero derivatives at both ends, so the smootherstep's zero first and second derivative survive at the start of the fade as well as at its death. It is also cheaper than pow.

  falloff values away from 1 cost slightly less transition width than they used to, so fadeRoom and the plateau ceiling are a little more generous.

0.4.2

• The kernel only walls off the hugged edge now. It was masked to the widget's rectangle on all four sides and renormalized, so a pixel in the middle of a band could not sample the page below it: the average leaned toward the edge, the band came out lighter than the untouched pixels just outside it, and the difference landed on a straight line at the far side — the exact artifact the widget exists to avoid. The one-sided kernel also dragged nearby content along the pass direction, which read as streaks. Only the hugged edge is a real boundary; the other three sample freely.

0.4.1

Two ways a band could still show a visible edge, and the number that prevents the second one.

• blurPlateau is now bound by the same ceiling as plateau. It was going to the shader raw, so a caller could hold the sigma over more of the rectangle than the fade had room for — the one combination of knobs left that produced a block with a findable edge. Haze.resolve scales both, and constrains the sigma with whichever holds longest. It returns a third value.

• plateau: 1 is now legal, and is the value to reach for. The assert rejecting it dated from when the value went to the shader raw, where 1 meant no fade at all; it is read as a fraction of the affordable room now, so 1 is the safest value there is rather than the worst.

• Haze.fadeRoom(sigma, [falloff]), the span a band needs BEYOND the chrome it covers. Too short a rectangle is the one thing the widget cannot fix for you: it keeps the effect smooth by shortening the plateau, which leaves your bar sitting on the ramp — a bar you can see through. Size the rectangle at chrome + fadeRoom(sigma), pass plateau: 1, and the plateau lands exactly on the chrome's edge.

0.4.0

The tint moved into the shader, and gained the option of reading the backdrop.

• The wash is no longer a LinearGradient. It was 30 piecewise-linear segments approximating a smootherstep, and a piecewise-linear approximation of a curve breaks its second derivative at every stop — which the eye reads as faint lines, one where the plateau gives way and one where the fade lands. It is now evaluated per pixel in haze.frag, on the same profile as the sigma, with a sub-LSB dither so 8-bit quantisation cannot band it either. The gradient remains for the pre-Impeller fallback, which has no shader to put it in.
• New: blurPlateau. The sigma's own plateau, independent of the scrim's. Null (the default) keeps them on one plateau, as before. They are separate because they are shaped by different things and measured differently: a flat band of wash has an end you can point at in a single screenshot, a blur's takes tracking edges across frames. It may be longer than plateau — measurement says it usually is — since blurCurve then makes the blur die first anyway. Whatever it is set to, the profile still reaches zero inside the rectangle, so it cannot produce a hard edge.
• New: tintAdaptivity. 0 (default) keeps the flat wash. 1 derives the scrim's alpha from the backdrop's luminance, so a white page is covered far less than a mid-grey one — iOS 26's behaviour, fitted to measurements of the system effect (the samples are in haze.frag). The gradient fallback — used during route transitions and pre-Impeller, where there is no backdrop to read — applies the same law at an assumed mid-range distance rather than ignoring it, which would paint the tint solid.

0.3.0

Everything in this release is one idea: a progressive blur must never have a findable edge, and that is the package's job rather than the caller's.

Breaking

• extent is gone. The ramp always spans the whole rectangle; to shorten it, shorten the rectangle. It existed as a dead margin against a faint line at the boundary, which the fixes below remove the cause of.
• falloff now warps the position along the ramp instead of the result. Same direction and the same meaning, 1.0 is unchanged, but values either side of 1 shape the middle of the curve slightly differently.
• blurCurve defaults to 2, so the blur sheds its radius earlier than in 0.2.0. Pass 1 for the previous behaviour.

Fixed

• falloff below 1 ended the blur on a vertical drop. pow(s, p) has an infinite derivative as s reaches 0 for any p < 1, which destroyed the smootherstep's flat landing — the effect stopped on a line and read as a plain BackdropFilter in a ClipRect. The exponent now applies to the position, leaving the smootherstep as the last link, so the ramp lands flat for every value.
• Haze threw under unbounded constraints. The blur render object is sizedByParent, and the internal stack expanded, so a Haze asked to be as tall as its child asked the shader for an infinite height. The child now sizes the widget.
• A large plateau or a falloff far from 1 could squeeze the fade into fewer pixels than the blur spans, which is a slab with an edge however smooth the curve. The transition (profile 0.9 down to 0.1) is now kept at least as wide as 3 × sigma: plateau yields first and proportionally, sigma only when even a plateau-free rectangle is too short. See Haze.resolve.

Added

• blurCurve, a perceptual exponent on the blur alone. Apparent blurriness tracks 1 / sigma, so a ramp linear in sigma hides its whole visible change in the last stretch. This is the sigma shaping 0.2.0 rejected — it was rejected because it was tried on the shared profile, where it also flattened the tint.
• plateau now holds the blur as well as the tint. In 0.2.0 the wash sat flat over the bar while the radius was already shedding underneath it, so the boundary the tint hid was the one the blur drew.
• Haze.falloffAt, Haze.resolve, Haze.plateauCeiling, Haze.sigmaCeiling and Haze.transitionFraction are public — to line something else up with the fade, or to show what the rectangle actually allowed.
• example/ rebuilt as a one-page studio: live sliders, the profile curve plotted from Haze.falloffAt itself, the resolved values when the rectangle overrules you, and a one-tap switch to a plain BackdropFilter for contrast.

0.2.0

• plateau: the tint can be held at full strength before it starts to fade, which is the shape of iOS's scroll edge effect — constant over the bar, then a fade whose start is as undetectable as its end (smootherstep, zero first and second derivatives at both ends). More gradient stops with it, so the gradient's own quantisation cannot reintroduce the banding the curve exists to remove.

  Tint only; the blur was byte-for-byte 0.1.1. (0.3.0 puts the blur on the same plateau — the two disagreeing was a bug, not a finding.)

0.1.0

• Initial release: Haze, a shader-based progressive blur with a matching tint wash, on any of the four edges.