2026-03-16
Peripheral Color: From Cliff to Gradient
Foveated renderers tend to desaturate the periphery too aggressively. The assumption is intuitive (acuity drops, so color must too), but the biology says otherwise. All three cone classes persist across the retina. L and M cones tile the periphery at lower density; S-cones were never dense to begin with, so their relative drop is modest. Psychophysical measurement puts blue-yellow sensitivity at 15° at ~79% of its 5° value, red-green at ~29% (Bowers et al. 2025). The periphery sees color, at lower spatial resolution.
Scrutinizer’s previous color pipeline missed this. A single
smoothstep ramp compressed the entire transition into one degree:
full color at the fovea, gray by 2°. A red banner and red body text at the same
eccentricity got identical treatment.
| Original |  |
| Before |  |
| After |  |
Original: unfoveated. Before: color dies at 2°. After: gradient rolloff, blue-yellow survives further than red-green.
What changed
Wider frequency coverage. The Difference-of-Gaussians decomposition extended from 8 bands (down to 0.35 cpd) to 12 bands (down to 0.088 cpd). The old range covered text and icons. The new range covers panel backgrounds, banners, and page-level color, the scales where peripheral color perception is strongest, because S-cone signals pool spatially over large areas.
Swatch-aware preservation. The coarsest MIP level
(mip[12], averaging ~64×64 source pixels) provides a free
swatch signal. Its Oklab chrominance magnitude distinguishes large uniform color
regions from mixed content. Swatches retain up to 30% more color than text at the
same eccentricity.
Gradual ramp. The base desaturation ramp widened from 1°–2° to 1°–6°. The per-band CSF attenuation handles frequency-specific rolloff. The ramp now provides a gradual spatial envelope.
Calibrated RG decay. The red-green channel decay rate
(rg_decay) increased from 0.072 to 0.085. At 12.4° (ring 5), the
model’s BY/RG ratio is 1.86×, monotonically separating from 1.05×
at 2.5°. The single-exponential decay cannot fully capture the biphasic RG
falloff that
Bowers et al. (2025) measured
(steep to ~15°, then slowing), but the rank ordering and channel separation
are correct across the measured range.
Validation
24px colored dots at five eccentricity rings (2.5°–12.4°), ~0.94 cpd at 45 ppd. The table gives the analytical model’s retention from the 2026-03-17 validation run; rendered screenshots are checked against it below:
| Color | 2.5° | 7.5° | 12.4° |
|---|---|---|---|
| Red | 82.5% | 57.8% | 39.4% |
| Green | 85.5% | 65.2% | 49.6% |
| Blue | 94.6% | 84.8% | 73.5% |
| Yellow | 94.2% | 83.9% | 72.5% |
Tier 1 (must pass) 9/9
All four colors decrease monotonically. BY retention ≥ 1.5× RG at
12.4° (1.86×). Both model predictions and rendered screenshots pass.
Tier 2 (should pass) 2/3
BY/RG channel ratio at 12.4°: model 2.59 vs
Bowers 2.72 at 15°
(5% off, threshold 20%). Green tracks red more than blue (10pp vs 24pp gap).
Both pass. The third check fails because rendered pixel values match the analytical model within
15% for only 5/20 measurements. The MIP chain and base desaturation ramp
introduce quantization the model doesn’t predict. Rank ordering is
correct; absolute values diverge.
Tier 3 (stretch) 3/3
Model retention correlates with
Hansen et al. (2009) color naming
accuracy at r = 1.000. BY outranks RG at every eccentricity ring (20/20).
Update (2026-10-02): The Hansen et al. (2009) color naming accuracy used in the Tier 3 check did not come from that paper, which reports chromatic detection thresholds at 10–50°. The data file and the check have been removed. The check was a rank correlation of two curves that both fall with eccentricity, so it returned r = 1.000 for any monotone model and could not fail. Tier 3 is now the single BY-above-RG check (20/20).