The current color matching uses a polar LCH decomposition with weights WL=0.5, WC=3.0, WH=6.0. Near-achromatic palette colors (black, white) have near-zero chroma, which collapses the hue difference term ΔH² to zero regardless of the pixel's hue. This makes black a universal attractor for saturated colors — vivid purple maps to black (dist=0.38) instead of blue (dist=1.28) because the hue penalty vanishes.
Additionally, the low lightness weight (WL=0.5) deprioritises structural detail, which human vision relies on for spatial recognition.
Change: Replace the LCH distance with weighted Euclidean distance in Cartesian OKLab:
dist = (1.0·ΔL)² + (1.5·Δa)² + (1.5·Δb)²
The Cartesian formulation has no achromatic singularity — the distance from a saturated color to black is naturally large because Δa² and Δb² remain intact rather than being subtracted out by −ΔC². The modest Wab=1.5 boost ensures e-paper color inks are still utilised without washing out neutrals.
Validation: Optimal Wab=1.5 determined by downsampled ΔE sweep (4×4 block averaging in OKLab) across 8 test images. No regressions on natural photos; measurable improvement on saturated out-of-gamut subjects.
Implementation:
claudobahn@014a00f
Before/After:
The current color matching uses a polar LCH decomposition with weights WL=0.5, WC=3.0, WH=6.0. Near-achromatic palette colors (black, white) have near-zero chroma, which collapses the hue difference term ΔH² to zero regardless of the pixel's hue. This makes black a universal attractor for saturated colors — vivid purple maps to black (dist=0.38) instead of blue (dist=1.28) because the hue penalty vanishes.
Additionally, the low lightness weight (WL=0.5) deprioritises structural detail, which human vision relies on for spatial recognition.
Change: Replace the LCH distance with weighted Euclidean distance in Cartesian OKLab:
The Cartesian formulation has no achromatic singularity — the distance from a saturated color to black is naturally large because Δa² and Δb² remain intact rather than being subtracted out by −ΔC². The modest Wab=1.5 boost ensures e-paper color inks are still utilised without washing out neutrals.
Validation: Optimal Wab=1.5 determined by downsampled ΔE sweep (4×4 block averaging in OKLab) across 8 test images. No regressions on natural photos; measurable improvement on saturated out-of-gamut subjects.
Implementation:
claudobahn@014a00f
Before/After: