The sunlit cloud tops were clipping to pure white. This is the fix, shown honestly: drag each slider to wipe between before and after, with the measured numbers under each — and the proof it changes only what was broken.
Result: a luminance soft-knee on the lit crowns. At the worst vantage the white-clip goes 14.7% → 0%, the crowns get their lobed form back, brightness barely moves (mean 227 → 222), and every vantage that wasn't blowing out is byte-for-byte unchanged. It's behind a flag (?softknee=0.5) — the confirmed-best look is the default and stays untouched.
It is not a bug-fix — the clip is faithful to Bonkahe (his shader clips the same way). This is a deliberate, surgical art-choice to keep the crisp look you picked while stopping the bleed.
Frozen, identical cloud state — the only thing changing is the soft-knee. Drag to wipe.
The two vantages where it shows — the top-down and the crown-graze — looping, before vs after.
Blown-to-white pixels (luminance ≥ 250/255) and mean luminance, real GPU, frozen frames. The fix lands only where the bleed was; the rest doesn't move.
| Vantage | blown before | blown after | mean L before→after |
|---|
SunshineCloudsCompute.glsl:726 accumulates in-scatter raw — lightColor += pow(color·weight, 2.2)·powder — with no view-transmittance weighting, identical to our line 380; our one extra factor uLightGain is 1.0 (a no-op). So his Godot clips the lit crowns the same way. "Fix the math" was ruled out — it would diverge from him.?kneemode=1 keeps per-channel for A/B.Sound, faithful-when-off. Verified the ceiling math (0.85 < the 0.902 clip threshold), the preserved gradient, and that nothing below byte ~217 can move (shadows/midtones/sky pass unchanged). Two flags, both addressed: (1) per-channel hue shift → adopted the luminance form; (2) the knee caps crowns below the bloom threshold, so it can quiet crown silver-lining glow → verified a bloom-on frame: crowns keep form and a soft glow (the HDR sun disc still blooms). Suggested strength: 0.5 (more shoulder recovery) ↔ 0.7 (minimal touch) — the CD's eye sets it.
Net improvement, no downside — ship it. Across all six vantages: zero regression — no view loses form, shifts hue, bands, or dulls; midtones and shadows are untouched. It helps most in the bright top-down and deck views (de-clips the crowns, recovers tonal separation) and is a clean no-op where there's nothing to fix (fly-by, distance, inside). The only "cost" — slightly softer highlight sparkle in the tops — is fake contrast against the 255 clip wall (those pixels carried no detail), so removing it isn't a loss of real form. They also surfaced two unrelated issues (not the soft-knee's): the distance and fly-by views render as featureless washes — a separate thread worth its own look.
?softknee=0.5 crown soft-knee strength (0 = OFF = faithful default) ?softrange=0.35 rolloff range above the knee (ceiling = softknee + softrange, keep < 1.0) ?kneemode=1 per-channel instead of luminance (A/B)
e.g. …/?module=5&atmo=0&softknee=0.5 — try 0.5 vs 0.7 on the top-down view and a warm-lit one.
Real-GPU captures (NVIDIA RTX 4070 Ti, D3D11), luminance/clip measured per pixel in display sRGB (qa/tools/measure_brightness.py), frozen frames for matched A/B, ffmpeg GIFs. Verdict grounded in Bonkahe's own source. Evaluated by two neutral teams — math-checker + scene-impact — two-directional briefs, no poison. The fix is a flagged option; the default render is unchanged. The CD's eye sets final strength.