Experiment #4 (blue×0.95) made shadow regions appear MORE blue, not
less — reducing source blue still leaves positive blue error after a
BLACK mapping, and the dither spends that error on neighbours,
creating blue dither dots in dark regions.
Reverting blue_mul to 1.0. Experiment #5 takes a different attack on
the same problem: shift the BLUE palette mapping target from the
muted (24, 64, 192) to a more saturated (8, 32, 220). Doesn't change
what the panel displays (the blue ink is fixed); it just makes
Euclidean distance from skin tones to "BLUE" larger in the algorithm's
view, so the dither prefers RED/WHITE/YELLOW for borderline pixels.
Render-compare's BASELINE struct now carries its own frozen palette,
so half-A keeps the original (24,64,192) BLUE target while half-B
pulls the shifted palette from the live pipeline.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Experiment #3 (sat 115, gamma 1.2, blur 0.6) was a net loss — greens
desaturated, no help on sky→face blue bleed.
Reverting those three to baseline (130, 1.0, 0.0). New experiment #4:
multiply the source's blue channel by 0.95 before dither. Real sky
stays well above the BLUE-vs-WHITE boundary, but borderline-bluish
skin (sky cast on outdoor faces) drops below it and maps to YELLOW /
WHITE / RED instead of feeding the dither's BLUE attractor.
Adds public BLUE_CHANNEL_MUL constant; render-compare's baseline
struct gets blue_mul=1.0 so half-A is still frozen at the original.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Riemersma dither produced visible "ink-spill" Hilbert-curve streaks in
low-contrast regions like skin. Reverting DITHER_METHOD to Floyd-
Steinberg and attacking the original sky-bleeds-into-face problem with
a pre-dither Gaussian blur instead:
DITHER_METHOD RIEMERSMA → FLOYDSTEINBERG (back to baseline)
+ new BLUR_SIGMA = 0.6 (pre-dither softening; baseline 0.0)
Live tunables are now public so RenderCompareCommand can mirror them.
Half-A of app:render-compare is locked to the frozen-baseline set
(sat=130 gamma=1.0 sharpen=0.8 blur=0.0 FS) — what shipped at git tag
render-baseline-2026-05-14. Half-B always tracks the live pipeline, so
each future experiment is automatically the next A/B without touching
the comparison command.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Stacks two dither treatments of the same image's top half into one V2
portrait .bin — top half Floyd-Steinberg, bottom half Riemersma —
overwrites the device's current V2 portrait asset, bumps rendered_at
and clears the preview PNG cache.
Usage: bin/console app:render-compare <imageId>
Lets Matt eyeball both methods on a single panel refresh instead of
two re-renders and two waits. One-shot experimental tool; not part of
the live render pipeline.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
football2801