@woosh/meep-engine 3.60.4 → 3.60.6

package/src/shade/renderer/shader/chunk/brdf/FRESNEL_BRIEF_2026_10_16.md

@@ -1,391 +0,0 @@
-# Fresnel for the direct lobe and the split-sum LUT: a brief (2026-10-16)
-
-**Status: decided and built.** Alex chose Schlick, in the form Unreal's rasterizer uses; it landed in dbdb1cdb1
-(§7). The sections before §7 are the brief as it was written, with the shading it describes as current. Every number
-here was measured; how, and where the scripts are, is in §8.
-
-Shade evaluates one material with two Fresnels. The direct lobe (`chunk_BRDF_GGX`) uses `F_Hauber`; the split-sum LUT
-`tSplitSum`, read by every indirect specular path and by `re_direct_multiscatter`, encodes `f0 · A + f90 · B` with
-Schlick's form at the fourth power. `MULTISCATTER.md` records three consequences: a dielectric's indirect specular is up
-to 0.068 of a uniform sky above its direct, a white metal's direct path reaches only 0.93, and Turquin's eq. 16
-overshoots the Heitz 2016 random walk partly because Hauber's dip compounds on every bounce.
-
-## 1. Recommendation
-
-**Use Schlick's `(1 − VoH)⁵` in the direct lobe, and regenerate the split-sum LUT with the same weight in both
-channels.** Keep the LUT, a 64² rg16float texture read with one fetch, and replace its contents. In §5 this is
-pairing E.
-
-- **All three inconsistencies close.** Direct and indirect agree to the LUT's precision: the worst gap in the f64
-  model goes from 0.071 to 0.000. A white metal's direct furnace goes from 0.929 to 1.000. At roughness 0.99, eq. 16
-  against a random walk using the same Fresnel goes from +4.7% to 0.0% for a white metal and from +5.7% to +1.0% for
-  silver.
-- **It costs one multiply per light.** Per vec3, it is 13 DXIL float operations against Hauber's 12, with no
-  transcendental. The LUT fetch and the per-pixel compensation are unchanged.
-- **Against exact Fresnel it ties with Hauber.**
-  - Dielectrics: the two are within 0.0033 of each other at every angle.
-  - Five measured metals: mean |ΔF| 0.0139 against Hauber's 0.0148. Each wins on different metals (§3).
-  - It is far more accurate than what the indirect path uses now. The shipped x⁴ LUT is up to +0.079 in albedo at
-    IOR 1.5 (up to +29%).
-- **What it changes visually** (§6):
-  - Dielectric reflections at grazing view darken toward the exact value: −12% to −34% of a sphere rim's linear
-    luminance. At roughness 0.26, N·V 0.26, the indirect albedo goes from 0.288 to 0.2225, and exact is 0.229.
-  - Metal spheres move by less than 2.3% at the rim.
-  - A back-lit white-metal or silver rim brightens by about 1%.
-- **It is the white-tint case of F82-tint**, the metal Fresnel of OpenPBR and Adobe's Standard Material. If a metal
-  edge tint is ever added to the material, it becomes a third LUT channel on top of this, not a rewrite (§4).
-
-**What would reverse it: preferring the grazing dip of high-index grey metals over silver, gold, copper and the perfect
-conductor.** For aluminium and iron at low roughness, Hauber is 0.03 to 0.06 closer in albedo; chromium, nickel and
-titanium follow the same physics. For silver, Hauber is 0.06 worse.
-
-If that look matters, Hauber's consistent form is pairing B (§5): it adds an `Ess` channel to the LUT. That closes the
-gap, but the white furnace stays at 0.93 and the walk overshoot at +4.7%.
-
-## 2. Why the two paths disagree: Hauber's dip forces a choice
-
-At `f90 = 1`, which every caller passes, `F_Hauber = f0 (1 − x⁴) + x⁵` with `x = 1 − VoH`. At `F0 = 1` it is
-`1 − VoH x⁴`, which dips to 0.918 at VoH = 0.2.
-
-Turquin's compensation needs `Ess`, the albedo with `F ≡ 1`, and Shade reads it as `A + B`. A two-channel LUT can hold
-Hauber's split or `A + B = Ess`, but not both:
-
-- **Before 933feac74**, the asset held Hauber's split exactly. That commit's own check found A within 4.9e-4 of the
-  `1 − f⁴` term and B within 4.8e-4 of the `f⁵` term. Its `A + B` was therefore the dipped white albedo. 933feac74 read
-  that shortfall (up to −8.2%) as a stale asset.
-- **933feac74 moved B to x⁴.** Now `A + B = Ess`, but `f0 · A + B` is Schlick at the fourth power, not Hauber, and the
-  indirect path has used a different Fresnel since. The furnace's 0.068 dielectric gap is that difference:
-  `(1 − f0)x⁴ − (x − f0)x⁴ = VoH x⁴`, up to 0.082.
-
-**The f64 model of the pre-933feac74 pairing gives no gap and a white furnace of 1.000, but for the wrong reason.**
-`Kms = (1 − A − B) / (A + B)` counts the dip as missing multiple scattering, so it stays positive even as roughness
-goes to 0.
-
-**Any Fresnel that is linear in F0, with `F(1, μ) ≡ 1`, removes the choice.** `f0 · A + f90 · B` is then its split,
-and `A + B` is its `Ess`. Schlick at any exponent qualifies; Hauber does not.
-
-## 3. Hauber's approximation, re-examined
-
-### Source
-
-Niklas Hauber, "Finding Alternatives to Schlick's Approximation of the Fresnel Equations", photometric.io, 2024-06-27
-(the URL in `chunk_F_Hauber`). He found it by GPU symbolic regression, with parameters R0 and 1 − cos θ:
-
-    R(θ) = R0 + (1 − cos θ − R0)(1 − cos θ)⁴
-
-**What he fitted it to:**
-
-- 4,096 points, θ uniform "from 0 to π" as written.
-- 40% at IOR 1.5, and 40% at an IOR uniform in [1.34, 1.7].
-- 20% from seven complex IORs: 0.3+3i, 0.15+4i, 0.3+3.35i, 1.4+7i, 1.84+3.73i, 2.1+4i, 2.9+3.1i.
-
-So the fit covers both dielectrics and conductors, mostly dielectrics, and most of its metals have a real part large
-enough to dip.
-
-**What he reports:** 34% lower MSE than Schlick, essentially the same shape for dielectrics, almost 50% better on his
-metals, with the warning that another dataset would differ. He also reports 5 CUDA instructions against 6. He offers a
-shape blend, `R0 + (1 − R0 − s cos θ)(1 − cos θ)⁴`, where `s = 1` is his form and `s = 0` is Schlick at the fourth
-power. The post does not discuss F0 → 1.
-
-### Shade's form
-
-`mix(f0, f90 − VoH, x⁴)` is `f0 + (f90 − VoH − f0) x⁴`. At `f90 = 1` this is Hauber's formula exactly, and it equals
-Schlick x⁵ minus `F0 · VoH x⁴`: the dip is proportional to F0, so it is deepest at F0 = 1. The `f90` is
-Shade's own generalisation; the post has none. The dip term does not scale with it, so at `f90 = 0` the form is
-negative (`−VoH x⁴`). Every caller passes 1, so that never runs.
-
-**His claim reproduces on his dataset:** 35.5% lower MSE (θ in [0, π/2]; the same with |cos θ| over [0, π]), 1.6% on
-dielectrics, 49% on metals. So the formula is read correctly.
-
-### Against exact Fresnel
-
-**Dielectrics** use the exact unpolarised Fresnel. Error over θ uniform in [0°, 90°]; "albedo" is the GGX lobe's
-directional albedo at the 25 furnace points (roughness 0.26–0.99, N·V 0.10–0.99):
-
-| IOR 1.5 (F0 0.04) | max \|ΔF\| | mean \|ΔF\| | ΔF_avg | worst Δalbedo | worst relative Δalbedo |
-|---|---|---|---|---|---|
-| Hauber | 0.033 at 85° | 0.0093 | −0.0068 | −0.0185 | −21.5% |
-| Schlick x⁵ | 0.036 at 85° | 0.0092 | −0.0061 | +0.0194 | −20.2% |
-| Schlick x⁴ (the shipped LUT) | 0.103 at 82° | 0.0214 | +0.0122 | +0.0785 | +28.7% |
-
-The worst absolute and the worst relative errors fall at different points. For Hauber and Schlick x⁵ the worst
-relative error is at mid angles, where the albedo is small and both read low (0.054 against 0.069 at roughness 0.51,
-N·V 0.51).
-
-Over IOR 1.3–2.0, the worst case for each:
-
-| | max \|ΔF\| | mean \|ΔF\| | worst \|Δalbedo\| | worst relative |
-|---|---|---|---|---|
-| Hauber | 0.067 | 0.0119 | 0.047 | −21.5% |
-| Schlick x⁵ | 0.068 | 0.0121 | 0.048 | −20.2% |
-| Schlick x⁴ | 0.142 | 0.0313 | 0.109 | +54.1% (IOR 1.3) |
-
-For a dielectric, Hauber is Schlick x⁵ minus `F0 · VoH x⁴`, at most 0.0033 at F0 0.04.
-
-**Conductors** use measured n and k: Johnson & Christy 1972 for gold, silver and copper; Johnson & Christy 1974 for
-iron; McPeak et al. 2015 for aluminium (Rakić 1995 has 13 rows in the visible, and its dip agrees to 0.012). The full
-complex Fresnel equations are integrated over 380–780 nm against D65 and the CIE 1931 2° observer (the CIE's own
-1 nm tables, taken at 5 nm), then projected to linear sRGB. That gives gold F0 (1.038, 0.729, 0.365) and aluminium
-(0.906, 0.916, 0.923). Each candidate is fed the exact curve's own F0; gold's red is left above 1.
-
-Worst channel per metal. The Δalbedo column is the worst over the 25 lobes:
-
-| metal | Hauber max \|ΔF\| | Schlick x⁵ max \|ΔF\| | Hauber Δalbedo | Schlick x⁵ Δalbedo | measured dip below F0 |
-|---|---|---|---|---|---|
-| silver | 0.078 | 0.007 | −0.065 | +0.005 | 0.3% |
-| gold | 0.062 | 0.033 | −0.051 | +0.029 | 4.5% (red) |
-| copper | 0.055 | 0.043 | −0.044 | +0.033 | 1.3% (red) |
-| aluminium | 0.049 | 0.109 | −0.026 | +0.082 | 7.1% (red) |
-| iron | 0.141 | 0.178 | +0.103 | +0.135 | 4.7% (red) |
-| mean of the 15 channels' mean \|ΔF\| | 0.0148 | 0.0139 | | | |
-
-Over all five metals, Hauber's MSE is 22% below Schlick's. That comes from aluminium (85% lower) and iron (47%). On
-gold it is 5× Schlick's, on silver 212×, on copper about equal.
-
-**As F0 → 1**, Hauber's minimum over angle falls further below F0:
-
-| F0 | 0.5 | 0.8 | 0.9 | 0.95 | 0.98 | 1.0 |
-|---|---|---|---|---|---|---|
-| Hauber, minimum | −0.5% | −3.4% | −5.4% | −6.7% | −7.6% | −8.2% (0.918 at VoH 0.20) |
-
-Schlick at either exponent never falls below F0, and is exactly 1 at F0 = 1. Only a perfect conductor has F0 = 1, and
-it has no dip. Silver, the nearest real metal (F0 0.98), dips 0.3%.
-
-### The dip belongs to n, not to F0
-
-Fitting Hauber's shape parameter `s` to each measured channel by least squares:
-
-| | silver | gold | copper | aluminium | iron |
-|---|---|---|---|---|---|
-| best `s` (R, G, B) | 0.09, 0.09, 0.05 | 0.38, 0.24, 0.58 | 0.34, 0.86, 0.88 | 1.15, 0.88, 0.64 | 2.24, 2.21, 1.96 |
-| F0 (R, G, B) | 0.99, 0.98, 0.98 | 1.04, 0.73, 0.36 | 0.93, 0.62, 0.52 | 0.91, 0.92, 0.92 | 0.53, 0.51, 0.49 |
-
-Silver and aluminium have nearly the same F0 and need opposite shapes. The dip comes from the real part of the index
-(Lazányi's correction has its slope at 2n), and the G-buffer carries only F0.
-
-**So no function of F0 alone can carry the dip.** Hauber's fit applies the average dip of its seven metals to every
-F0, including F0 = 1. It is right for aluminium-like metals and wrong for silver-like ones.
-
-## 4. The state of the art in pure functions
-
-### Fresnel
-
-DXIL is from Chrome's own Tint and DXC on the RTX 4090. Float operations per vec3 are counted, with the input loads
-excluded.
-
-| form | inputs | DXIL float ops (div, sqrt, exp/log) | five metals: worst max \|ΔF\| / mean \|ΔF\| | dielectric IOR 1.5 max \|ΔF\| | F0 = 1 |
-|---|---|---|---|---|---|
-| Schlick 1994, x⁵ (multiplies) | F0 | 13 (0) | 0.178 / 0.0139 | 0.036 | 1 |
-| Schlick x⁵ as `chunk_F_Schlick` (`pow`) | F0 | 13 (log + exp) | same | same | 1 |
-| Schlick x⁴ (the shipped LUT) | F0 | 12 (0) | 0.213 / 0.0186 | 0.103 | 1 |
-| Hauber 2024 (shipped direct) | F0 | 12 (0) | 0.141 / 0.0148 | 0.033 | 0.918 |
-| Lazányi–Schlick 2005 | n, k → F0, a, α | 25 (log + exp) | needs n | — | — |
-| F82-tint: Hoffman 2019, Adobe ASM, OpenPBR | F0 + edge tint | 40 (3 div) | 0.020 / 0.0026 (tint from data) | — | 1 at tint 1 |
-| Gulbrandsen 2014, (r, g) → n, k | F0 + edge tint | 109 (15 div, 3 sqrt) | exact up to its approximate conductor formula | — | singular, clamped to 0.99 |
-| exact dielectric (`F_dielectric`) | IOR | 28, scalar (4 div, 1 sqrt), branch | — | 0 | — |
-| exact conductor | n, k | 85 (6 div, 6 sqrt) | 0 | — | — |
-
-Notes on the sources (the literature summaries are with the scripts, §8):
-
-- **Schlick 1994** was not read first-hand: the only reachable copy is an image-only scan. Its formula is as Hoffman
-  2019 (eq. 1) and OpenPBR give it.
-
-- **Lazányi & Szirmay-Kalos 2005** (WSCG): subtract `a μ (1 − μ)^α` from Schlick. `a = 2n` comes from the slope at
-  grazing incidence; α is fitted at μ = 0.15. Their stated relative errors: Cu 1.9–4.9%, Au 5.0–5.4%, Ag 4.4–9.0%,
-  Al 2.0–17.5%. The inputs are n and k.
-- **Gulbrandsen 2014** (JCGT 3(4)): maps reflectivity and edge tint to n, k, then evaluates PBRT's approximate
-  conductor formula. Standard Surface uses it; MaterialX feeds the result to its full conductor Fresnel.
-- **Hoffman 2019 and 2023, Kutz et al. 2021/2023 (Adobe Standard Material), OpenPBR 1.1.1:**
-  `F = Schlick(μ) − b μ (1 − μ)⁶`, with `b = Schlick(1/7)(1 − tint)/(μ̄(1 − μ̄)⁶)` and μ̄ = 1/7, so that F(82°) =
-  tint · Schlick(82°). It is linear in F0, F90 and b, so a split-sum LUT needs one more channel, `∫ μ(1 − μ)⁶`.
-  The OpenPBR companion paper gives its hemispherical average as `F0 + (1 − F0)/21 − b/126`. Unreal 5.4 has the
-  Adobe F82 tint only in its Substrate path tracer (disabled by default); its rasterizer runs Schlick.
-- **Filament and Frostbite** use Schlick with a shadowed F90. Belcour et al. 2020 needs a tabulated basis.
-  Vernooij 2024 fits the Lazányi form per metal. No F0-only metal approximation newer than Hauber's was found.
-
-### Split-sum (DFG) fits
-
-**Keep the LUT.** It is one fetch that both paths already pay, since the compensation reads it anyway. It is also an
-order of magnitude more accurate than the best published fit, which costs 26 to 89 float operations.
-
-Each fit was measured against the f64 x⁵ split of Shade's lobe (GGX, height-correlated Smith, α = r²). The points are
-a 32² grid of N·V ≥ 0.05 and roughness, which falls midway between the 64-wide LUT's texel centres, its worst place to
-interpolate. DXIL is counted as above.
-
-| fit | fitted to | DXIL float ops | white (F0 1) albedo: max \|ΔE\| / mean | F0 0.04: max \|ΔE\| |
-|---|---|---|---|---|
-| **the LUT, regenerated at x⁵** | Shade's lobe | **1 sample** | **0.0027 (0.3%) / 0.0002** | **0.0009** |
-| Karis 2014 `EnvBRDFApprox` (`DFGApprox`) | UE4's G, "adapted" | 16 (1 exp) | 0.342 (−43%) / 0.106 | 0.083 |
-| Narkowicz 2014, polynomial | uncorrelated Smith, α = (1 − gloss)⁴ | 36 (1 sqrt) | 0.221 (−28%) / 0.019 | 0.080 |
-| Narkowicz 2014, Lazarov form | same | — | 0.315 (−39%) / 0.041 | 0.080 |
-| Pesce & Iwanicki 2015 | not stated; roughness = √r, the best of three readings | 15 (1 sqrt, 1 exp) | 0.126 (−15%) / 0.028 | 0.135 |
-| UE 5.4 analytic energy fit ("WIP") | height-correlated table | 26 (2 div, 3 log, 3 exp) | 0.058 (+6.5%) / 0.0064 | 0.015 |
-| MaterialX `mx_ggx_dir_albedo_analytic` (2022) | height-correlated, Schlick | 89 (2 div) | 0.084 (−8.4%) / 0.0060 | 0.025 |
-| Kulla & Conty 2017, E only | height-correlated | 38 (div, sqrt, log, exp) | 0.011 (3.5%) / 0.0022 | — |
-| Sforza & Pellacini 2022, E only | uncorrelated Smith | — | 0.226 (−28%) / 0.016 | — |
-
-Notes on the table:
-
-- **Worst cases.** Most fits are worst at grazing (N·V 0.08) and rough. The LUT's three columns below N·V 0.05 reach
-  0.019.
-- **Fdez-Agüera 2019 is not a fit.** It reads Epic's LUT, and its contribution is a multiple-scattering form, with
-  `Favg = F0 + (1 − F0)/21` for Schlick.
-- **Kulla & Conty keep tables.** Their slides favour tables over fits. The E-only fits can stand in for Turquin's
-  `Ess`, but not for the split.
-- **Turquin lists analytic fits as future work.** His conclusion names good analytic fits for all the smooth tables as
-  future work, to save texture fetches in real time. No published fit reaches the LUT's accuracy for a
-  height-correlated lobe.
-- **Engines moved from fits to LUTs in 2025.** three.js dropped Karis's fit for a height-correlated LUT (PR #32054,
-  October 2025). Godot's Forward+ renderer moved to a LUT in March 2025; its mobile renderers keep Karis's fit.
-  Filament, EEVEE and Unity HDRP use LUTs only.
-- **One Shade pass still uses a fit:** `ssr_combine_shader.js` weights by `DFGApprox`, which is 0.34 (43%) below Shade's
-  white albedo at grazing, rough. That is already reported (2026-10-03) and separate from this brief.
-
-## 5. The pairings against the three inconsistencies
-
-Measured in an f64 model of Shade's two paths under a uniform sky, at the 25 furnace points, for white, a 0.04
-dielectric and the five measured metals:
-
-- direct = `E_F(F0) · (1 + F0 Kms)`
-- indirect = `(F0 A + B)(1 + F0 Kms)`
-
-The model reproduces the real passes' furnace sweep (`white_furnace.spec.js`'s helpers, run through the emulator):
-white direct 0.9301 against 0.9302, dielectric 0.2198 against 0.2196, gap 0.0682 against 0.0684.
-
-The walk is the shared bench's port of Heitz 2016, 4M paths per point. Each Fresnel candidate and the exact Fresnel
-are applied at every bounce, and the conductors are walked per wavelength (10 nm) before projection.
-
-| pairing | worst indirect − direct | white direct, min | eq. 16 vs its own walk, r 0.99: white / silver | LUT |
-|---|---|---|---|---|
-| A. shipped: Hauber direct, LUT x⁴ | +0.071 (white), +0.068 (dielectric) | 0.929 | +4.7% / +5.7% | as is |
-| B. Hauber direct, LUT (1 − x⁴, x⁵) + an `Ess` channel | 0.000 | 0.929 | +4.7% / +5.7% | 3 channels |
-| C. Hauber direct, LUT (1 − x⁴, x⁵), `Kms` from A + B (pre-933feac74) | 0.000 | 1.000, by booking the dip as multiple scattering | — | regenerate |
-| D. Schlick x⁴ in both | 0.000 | 1.000 | 0.0% / +0.9% | as is |
-| **E. Schlick x⁵ in both** | **0.000** | **1.000** | **0.0% / +1.0%** | **regenerate** |
-
-**D needs no new asset, but it moves the direct path to the least accurate form.** For a dielectric, it puts the
-+29% albedo error of today's indirect path into the direct one, and it is worse on every metal (§3).
-
-**E is the shipped LUT's two channels, regenerated.**
-
-**Against the walk with exact Fresnel**, the worst over N·V and channels, at roughness 0.26:
-
-| | white | silver | gold | copper | aluminium | iron | dielectric IOR 1.5 |
-|---|---|---|---|---|---|---|---|
-| Hauber (A's, B's direct) | −7.1% | −6.9% | −8.0% | −5.4% | +3.4% | +17.3% | −22% |
-| Schlick x⁵ (E) | 0.0% | +0.5% | −4.1% | +4.3% | +10.7% | +23.4% | −21% |
-| Schlick x⁴ (A's indirect, D) | 0.0% | +0.6% | +6.4% | +9.2% | +11.5% | +28.9% | +25% |
-
-The dielectric's relative errors are on small values. At N·V 0.51 they are 0.0687 (x⁵) and 0.0962 (x⁴) against an
-exact 0.0869. At N·V 0.26 they are 0.2225 and 0.2879 against 0.2294.
-
-**At roughness 0.99, the Fresnel no longer decides; Turquin's `Fms = F0` does.** Against the exact walk, eq. 16 is
-+21.7% for iron, +17.2% for copper and +23.8% in gold's blue under every Fresnel. That is the compensation's own
-saturation error, which `MULTISCATTER.md` already ledgers.
-
-**Hauber's dip compounding is the part E removes.** Against walks with their own Fresnel at roughness 0.99, eq. 16's
-mean overshoot is:
-
-| | white | silver | aluminium | gold | copper | iron |
-|---|---|---|---|---|---|---|
-| Hauber | +0.026 | +0.029 | +0.036 | +0.024 | +0.031 | +0.025 |
-| Schlick x⁵ | 0.000 | +0.004 | +0.016 | +0.010 | +0.020 | +0.021 |
-
-Against the exact walk the totals move the other way, by 0.003 to 0.008 for the metals: Hauber's dip had been
-offsetting part of `Fms = F0`'s overshoot.
-
-## 6. The visual change
-
-Rendered on the RTX 4090 with the multiscatter bench's sphere sweep: seven roughnesses (0.1–1.0) by seven materials
-(gold, copper, aluminium, iron and silver at their table F0, a white metal at F0 1, and a black dielectric), IBL, TAA,
-EV −1.
-
-Two lightings:
-
-- the bench's sun, which comes from the camera's side, so VoH is about 0.9 on every sphere;
-- the same sun from behind the spheres, so their rims see VoH ≈ 0.27, where Hauber's dip lives.
-
-No engine file was edited. The "after" page's prelude rewrites `F_Hauber`'s body to Schlick x⁵ in the one module that
-declares it, and replaces `split_sum.bin`'s texels as they are fetched. The replacement is the shipped texels plus the
-f64 change of weight, `∫ VoH x⁴`, so only the Fresnel differs, not the integration. A repeat of the shipped arm
-differed by at most 0.27% (front) and 0.09% (back).
-
-Change in mean linear luminance, after against before, over the outer rim (0.75–0.92 of the radius):
-
-| | 0.1 | 0.25 | 0.4 | 0.55 | 0.7 | 0.85 | 1.0 |
-|---|---|---|---|---|---|---|---|
-| black dielectric, front | −34.1% | −32.8% | −27.6% | −21.4% | −18.4% | −14.9% | −12.3% |
-| black dielectric, back | −33.7% | −17.3% | −7.5% | −11.6% | −17.2% | −15.6% | −13.4% |
-| iron, front | −2.3% | −2.2% | −1.6% | −0.9% | −0.6% | −0.4% | −0.3% |
-| copper, front | −1.5% | −1.4% | −1.0% | −0.5% | −0.3% | −0.2% | −0.1% |
-| white metal, back | 0.0% | +0.9% | +1.3% | +0.8% | +0.2% | 0.0% | 0.0% |
-| silver, back | −0.1% | +0.6% | +1.0% | +0.5% | +0.1% | 0.0% | 0.0% |
-
-Over whole spheres, metals move by at most 0.4% and the black dielectric by 2–8%.
-
-**Where the eye sees it is the dielectric's Fresnel rim under the sky.** It loses the x⁴ LUT's excess and lands near
-exact: 0.2225 against 0.2294 at roughness 0.26, N·V 0.26, where it was 0.288.
-
-Swapping only the direct Fresnel to x⁴ (pairing D) brightens the back-lit black dielectric's rim by up to 28%. The
-images are with the scripts (§8).
-
-## 7. As built (dbdb1cdb1)
-
-**The form is standard Schlick.** `F_Schlick(F0, F90, VoH)` computes `Fc = (1 − VoH)^5` by multiplies, then returns `F90 Fc + (1 − Fc) F0`, and the split-sum lookup reads a LUT integrated with the same weight. `chunk_F_Schlick` now has that form; its `pow(x, 5.0)` had compiled to a log and an exp.
-
-**Every `F_Hauber` caller moved to it,** and `chunk_F_Hauber` is deleted:
-
-- `chunk_BRDF_GGX`;
-- `graph_lpv_shade.js`;
-- the path tracer: `chunk_render_trace_path.js` ×2 and `chunk_sample_material_data.js`;
-- the forward fragment;
-- the view Fresnel of `shader_avboit_splat.js` and `shader_oit_generate_moments.js`.
-
-**`chunk_integrate_brdf` weights `(1 − x⁵, x⁵)`, and `split_sum.bin` was regenerated once on the RTX 4090.** It
-followed the 16cca4c69 procedure: 2^18 samples in 1024-sample blocks through rg32float, f16 encoded on the CPU, then
-the chunk put back at 1024 samples. Checked against the old asset:
-
-- its A + B moved by at most one f16 ulp anywhere;
-- against the old asset plus the f64 change of weight, it is within 5.8e-4 everywhere.
-
-**Pinned red then green:**
-
-- `chunk_BRDF_GGX.spec.js` (new): a perfect conductor's lobe is D V alone, and the lobe's Fresnel is Schlick's fifth
-  power.
-- `split_sum_asset.spec.js`: weight x⁵, and the generator itself held to the same integral.
-- `white_furnace.spec.js`: the dip's slack goes from four checks.
-  - A white metal's direct light reads 0.997 – 1.000 of the sky.
-  - A pure material's direct and indirect specular agree to 0.0031.
-  - A white dielectric's totals agree to 0.0010.
-  - The metalness blend holds to 1.3e-5.
-- `chunk_F_Schlick.spec.js` (new): pins the form.
-
-**The diffuse coupling of 835cba76b now balances exactly.** It keeps 1 − the LUT's dielectric albedo, which is now the
-direct lobe's own, so a white dielectric's direct light reads 1.000 – 1.001 of the sky (it was 0.932 – 1.000).
-
-**Device checks:**
-
-- Every shader that links a changed chunk compiles on the 4090: 19 unique modules, found through the reverse import
-  closure. The lightmap bake pair fails on a struct member already missing on master (a224bc59c1).
-- The sphere sweep, rendered from a detached master worktree and from the change, moves as §6 predicted, to 0.1
-  point.
-
-**Not part of this:** `ssr_combine_shader.js` weights by Karis's fit, not the LUT (§4).
-
-## 8. Method and reproduction
-
-The scripts, data and renders are in the side-session notes folder
-`local_9644ea11-c179-40dd-a485-620475f78a45/fresnel-brief/`, outside the repo:
-
-- `dfg_lib.mjs`: the lobe integrated over microfacet normals, with GGX's CDF as the variable, so that narrow lobes
-  resolve. It is held to `split_sum_asset.spec.js`'s light-direction quadrature and to the shipped texels to about
-  1e-4.
-- `fresnel_lib.mjs`: exact dielectric and complex conductor Fresnel, and the spectral projection.
-- `optical_data.json`: n and k, CIE 1931 and D65, each with its source URL.
-- `study_dielectric.mjs`, `study_metal.mjs`, `hauber_claim.mjs`, `study_dfg.mjs`, `study_dfg_e.mjs`,
-  `pesce_mapping.mjs`, `combos_study.mjs` and `model_check.mjs`, each with its `.txt` output.
-- `dfg_fits.mjs`: every published fit, with its constants.
-- `lit_fresnel.md` and `lit_dfg.md`: the literature, read from the primary sources, with URLs and sections.
-- `walk_candidates.mjs`: the shared bench's Heitz port, with every Fresnel. Run with `analyse_walk.mjs`.
-- `alu/`: the dump page and counter.
-- `render_drive.mjs`, the sphere page and `sphere_diff.py`, with the renders.

package/src/shade/renderer/shader/chunk/color/tonemap/ACES_TONE_MAPPING_BRIEF_2026_10_17.md

@@ -1,153 +0,0 @@
-# ACES tone mapping, Unreal's default look: a brief (2026-10-17)
-
-**Status: built.** `ShadeToneMappingMode.ACES` is the default. It is added beside GT7, PBR Neutral and NONE, and
-replaces none of them. Alex asked for a close match to Unreal Engine's current default look, on SDR and HDR alike,
-as fixed-function tone mapping: no settings and no LUTs.
-
-## 1. What it is
-
-- **SDR is Unreal's "Filmic" tone mapper at its default settings.** It evaluates UE 5.4's sRGB LUT colour-grading and filmic tone-mapping pipeline in `chunk_tonemap_ue_filmic` per pixel where UE bakes it into a 32³ LUT:
-  1. ACEScg from UE's default sRGB working space;
-  2. Expand Gamut at 1.0;
-  3. the colour corrections at their defaults, which only clamp at zero;
-  4. Blue Correction at 0.6;
-  5. the curve at Slope 0.88, Toe 0.55, Shoulder 0.26, Black Clip 0 and White Clip 0.04 (the default post-process settings);
-  6. the inverse blue correction, then Rec.709 clipped to [0, 1].
-- **HDR is what Unreal 5.8 does with Filmic on a display past 100 cd/m²:** the SDR picture, taken back to the scene
-  through the inverse ACES 2.0 output transform for a 100 cd/m² display, and rendered again with the ACES 2.0 output
-  transform for this display's headroom, both limited to Rec.709. The ACES 2.0 part is OpenColorIO 2.6.0's
-  implementation, in `chunk_aces2_output_transform`, with three limits on its inverse (D5). Its per-display parameters
-  and gamut tables are built on the CPU in `aces2_transform_block.js` and owned by `TonemapAces2Tables`.
-
-| | SDR | HDR, 1000 cd/m² | HDR, 2000 cd/m² | HDR, 4000 cd/m² |
-|---|---|---|---|---|
-| exposed 0.18 shows at, × SDR white | 0.180 | 0.257 | 0.287 | 0.313 |
-| white reaches, × SDR white | 1.0 at +6 EV | 4.90 | 9.62 | 18.64 |
-
-HDR columns are 1015, 2030 and 4060 cd/m² displays, whose headroom over the 203 cd/m² paper white puts ACES 2.0 at
-500, 1000 and 2000 cd/m². The values are OpenColorIO's.
-
-Mid-grey rises with the display's peak. That is ACES 2.0's design, and Unreal 5.8 keeps it: 10 cd/m² at a 100 cd/m²
-display, plus 14% of that for every doubling of the peak. A display with no headroom shows the SDR picture exactly.
-
-## 2. Sources
-
-- **UE 5.4 documentation:** "Color Grading and the Filmic Tonemapper in Unreal Engine", Epic Games.
-- **UE after 5.4, release notes** (Epic, verified):
-  - 5.6 adds ACES 2.0 for HDR as an option.
-  - 5.8 adds "support for ACES 2.0 in SDR", moves ACES into a secondary LUT, and drops the HDR UI paper white from 300
-    to 203 cd/m² to match BT.2408.
-  - 5.8 offers Filmic (default) and Standard ACES tone-mapping methods.
-- **UE 5.8's behaviour, Epic staff** (Yutaro Sawada, forums.unrealengine.com/t/tonemappingmethod/2747218, 2026-07-30,
-  in Japanese):
-  - ACES 2.0 is now the default HDR version.
-  - Filmic, the default, aims at one look on SDR and HDR. Past 100 cd/m² it is "applied, converted back to linear
-    through the inverse ACES 2.0, then output again through ACES 2.0 at the target luminance".
-  - The scene colour multiplier applies to Standard ACES only.
-  - The peak comes from the hardware, and mid and minimum luminance derive from it.
-- **ACES 2.0:**
-  - ampas/aces-core, Lib.Academy.OutputTransform.ctl and Lib.Academy.Tonescale.ctl;
-  - OpenColorIO v2.6.0, ops/fixedfunction/ACES2/Transform.cpp, FixedFunctionOpCPU.cpp and transforms/builtins/ACES.cpp,
-    which this ports.
-- The glue between Filmic and ACES 2.0 (§3, D2) is inferred from the staff description.
-
-The state of the art elsewhere, for context:
-
-- Godot defaults to Linear, three.js to none, and Unity HDRP to none; Filament defaults to a legacy ACES fit.
-- Blender added ACES 1.3 and 2.0 views in 5.0, with AgX the default since 4.0.
-- No real-time engine but Unreal ships ACES 2.0 by default.
-
-## 3. Decisions
-
-- **D1. SDR is Filmic exactly, Unreal's pass excepted.** Left out because they are not tone mapping:
-  - vignette (UE's default intensity 0.4), grain, sharpening and bloom;
-  - the LUT itself: UE interpolates a 32³ table, where this evaluates the function;
-  - the LUT's input range: UE's log encoding stops at 43.7, so it shades anything brighter as 43.7. This does not,
-    which moves saturated highlights past 43.7 by a few codes.
-- **D2. HDR, the glue I inferred:**
-  - The inverse runs at 100 cd/m² with Rec.709 limiting primaries, because Filmic's output is a Rec.709 SDR picture.
-  - The forward runs at 100 × the display's peak over the paper white, also limited to Rec.709. So both transforms
-    take SDR white for 100 cd/m², and SDR white lands on the paper white, as in every other mode.
-  - Rec.709 rather than the canvas's Display P3, because the review measured what P3 does: just past the paper white
-    (203.01 cd/m²) saturated colours jumped up to 67 codes from the SDR picture, 213 of 512 by more than 2. Limited
-    to the SDR picture's own gamut, the HDR picture becomes the SDR one as the headroom shrinks to none. HDR gives up
-    colours outside Rec.709, which the SDR picture never had either.
-  - A display at or under the paper white shows the SDR picture, Unreal's "past 100 cd/m²" rule in this engine's
-    units.
-  - Unreal sets the HDR paper-white level in nits and runs the forward transform at the display's absolute
-    peak. If its paper white scales the SDR picture onto the paper white, which its default of 203, BT.2408 and ours
-    suggest, the two agree.
-- **D3. No scene colour multiplier.** UE applies its 1.5 to Standard ACES only.
-- **D4. Tables are built on the CPU, per display.** ACES 2.0 needs a reach table, a hue table and a gamut-cusp
-  table with an upper-hull gamma per hue. They are found by searches no pixel can afford.
-  - Built in 7–13 ms in JS, on the first HDR frame and then on a change of peak, not every frame and not with the
-    renderer.
-  - Bound as one 15 KB storage buffer: two blocks of 1,893 floats.
-  - These are the algorithm's own data, not a grading LUT.
-- **D5. The inverse is limited where OCIO's is ill-conditioned.** Filmic clips channels onto the Rec.709 cube's
-  surface, and ACES 2.0's estimate of the gamut boundary lies a hair inside that surface. So a bright tinted highlight,
-  SDR (0.981, 1, 1) say, sits on or past the boundary, where OCIO's inverse extrapolates. The review measured it as a
-  jump of 0.24 × SDR white between neighbouring scene values on a smooth ramp, which a picture shows as speckle or a
-  ring. Three limits, only in the inverse, and only where it is ill-conditioned:
-  - colourfulness past the estimated boundary is clamped to it, so it comes back as the reach boundary, which the
-    forward transform takes back to the boundary;
-  - the decompression's gain is held to 5. Below the last unit of J it stays under 4.9 for 3,000 Filmic outputs, and
-    within that unit it reaches 600 to 11,700, which multiplies float32 rounding by as much;
-  - over that last unit of J the decompression fades to none, so a near-white colour meets the neutral that white,
-    past J_max, is sent to.
-
-  A colour whose SDR picture is not in that region is untouched and matches OCIO. Two more departures are where OCIO
-  gives NaN: the gamut boundary's power takes a base clamped at zero, and an achromatic response at or under zero
-  tone-maps to zero, which is where OCIO's `max(0, NaN)` lands it.
-
-## 4. Measured
-
-**Against the references**, in `shader_tonemap_aces.spec.js`, under the WGSL emulator, in Display P3:
-
-- SDR is within the dither of a float64 Python reference of the Filmic curve and grading pipeline, at 29 colours from black to 1000.
-- HDR is within 1e-4 (encoded) of OpenColorIO 2.6.0 at four display peaks: its builtin inverse at 100 cd/m², then its
-  ACES 2.0 fixed function limited to Rec.709 between its builtins' clamps.
-  - The exceptions are three bright saturated colours whose SDR picture is near white, where D5's limits apply, and
-    a grey at +4 EV on 8120 cd/m², where OCIO's float32 inverse drifts by more than that. These four are held to a
-    float64 Python port with the same limits instead.
-- A tinted highlight's HDR picture moves continuously: the largest step along a ramp shrinks when the interval is cut
-  ten times finer. Without D5's clamp and gain limit it does not, nor with P3 limiting at 203.01 cd/m² is an HDR
-  display within 2 codes of SDR; both tests failed before the change.
-- Mutation testing: 110 mutants by an adversarial reviewer, then the survivors and every limit in D5 again. Three
-  survive, each unreachable or under the tolerance (the inverse's nd ≥ 1 branch, which the clamp makes unreachable;
-  the gain ≥ 1 escape; a looser cusp-search tolerance).
-- The JS table builder matches a float64 Python port to 6e-8.
-- That Python port matches OCIO's builtins to 2e-5, forward and inverse, at 100, 500 and 1000 cd/m².
-
-**On the device** (RTX 4090, headless Chrome, volumetrics_froxel, 1258×646, exposure fixed). OKLab lightness and HSV
-saturation are on the displayed frame:
-
-| view | mode | L mean | L p05 | L p95 | saturation |
-|---|---|---|---|---|---|
-| overview | ACES | 0.434 | 0.141 | 0.655 | 0.323 |
-| overview | GT7 | 0.445 | 0.180 | 0.635 | 0.281 |
-| broadside | ACES | 0.512 | 0.224 | 0.868 | 0.263 |
-| broadside | GT7 | 0.522 | 0.258 | 0.891 | 0.241 |
-| volume edge | ACES | 0.472 | 0.256 | 0.725 | 0.342 |
-| volume edge | GT7 | 0.480 | 0.288 | 0.699 | 0.300 |
-
-- Mid-grey is where GT7 puts it, 0.18 in and 0.18 out.
-- The mid-tone contrast is 1.36 against GT7's 1.05. Shadows sit deeper and highlights brighter, and saturation rises
-  12–15%.
-
-The pass's cost, three interleaved rounds of 24 frames:
-
-| pass | ACES | GT7 |
-|---|---|---|
-| SDR | 12–13 µs | 14–21 µs |
-| HDR | 55–102 µs | 16–33 µs |
-
-The HDR pass was forced through `matchMedia` and ran on the device with no validation errors and no non-finite output.
-It has not been looked at on an HDR display.
-
-## 5. Open, measured
-
-- **The HDR pass costs three to four times GT7's**: two ACES 2.0 transforms a pixel, 55–102 µs at 0.8 MP on a 4090.
-  Unreal pays this once into a LUT, which was ruled out here.
-- **The glue to Filmic is inferred** (D2, D5). Unreal evaluates this composite into a LUT, whose interpolation smooths
-  what D5 limits.