A misleading test can send a photo editor in the wrong direction. One almost gray photograph dominated Chiado’s color-combination error ranking because the diagnostic treated hue angles like ordinary numbers and gave unstable, nearly neutral colors too much influence.
We changed the diagnostic to use circular hue vectors weighted by the remaining chroma, then rechecked earlier rendering decisions with both metric versions. That corrected the priority list and preserved an audit trail. The score fell from 0.50723 to 0.03370 on the same images; the fix was to the development measurement, not to the pixels in that photograph.
A hue angle is not an ordinary number
Angles wrap. A hue turn of +179 degrees and a turn of −179 degrees are only two degrees apart around a circle, although ordinary subtraction puts them 358 degrees apart. Averaging angles as numbers creates a related problem: the arithmetic average of those two turns is zero, pointing in the opposite direction from either observation.
Our earlier diagnostic summarized pixel hue turns using an arithmetic mean, then compared the two renderers’ means with ordinary subtraction. That was convenient, and often behaved reasonably for small changes. Strong combined edits exposed its limits.
In the nearly gray case, the mean turns were about +120.68 degrees in Lightroom and −122.46 degrees in Chiado. The total diagnostic error was 0.50723; the hue component contributed about 0.48629. Those are units of our own measurement, not percentages or perceptual ratings.
Hue also becomes unstable as color disappears
Hue describes a direction in a color plane. Near its center, chroma is small: tiny changes in the two color coordinates can produce a large change in angle while the pixel still looks nearly neutral.
Simply replacing subtraction with the shortest angular distance would address wrapping, but it would still give too much importance to uncertain directions in almost gray output. The measurement needed to account for both geometry and remaining color.
Represent the change as a vector
The revised diagnostic represents a hue turn as a point on the unit circle, then subtracts the no-turn vector. A zero turn contributes zero. Adjacent directions stay adjacent across the wrap boundary.
Each pixel’s contribution is weighted by the smaller of its original and edited chroma, and the sum is normalized by the supported original chroma. As an edit removes a pixel’s color, that pixel’s hue contribution fades. We retain a baseline chroma threshold to exclude initially near-neutral pixels.
This is a particular engineering choice, not a universal perceptual formula. The vector distance is scaled to agree locally with our earlier small-angle weighting. It remains a summary: different spatial patterns or mixtures of turns can produce similar vectors. Visual inspection still matters.
A lower score did not mean a better photograph
Under the revised metric, the nearly gray case scored 0.03370 instead of 0.50723. The images were identical to the ones measured before. We had corrected the diagnostic, not improved the renderer.
The highest-scoring case then shifted to a strongly saturated, cool-graded image, where chroma and tone differences dominated. Inspection of the revised ranking still found real gaps: Lightroom could produce stronger cyan where Chiado tended toward purple or blue, and positive Saturation over reduced Vibrance could produce a stronger yellow/orange response.
The corrected metric helped us prioritize those differences. It did not make the residual warm/green cast in the nearly gray image disappear.
Keep the old ruler when introducing a new one
We preserved the original metric and its reports, gave the revised diagnostic a separate version, and recorded both values when rechecking prior calibration decisions. Recomputed legacy scores had to reproduce before we interpreted the new ones.
The positive-blue selection change retained its direction of improvement in every rechecked intermediate, adjacent-color and native-crop case. Native grading still improved in 17 of 18 cases, with the same small regression. The broader original grading sweep retained its aggregate benefit, but four individual improvement/regression classifications changed. A measurement revision therefore needs a case-by-case audit, even when the overall conclusion survives.
Those checks use photographs already consumed during development. They do not establish independent validation, other-camera performance or overall rendering quality. We also kept historical image precision unchanged in the grading rechecks so that we could isolate the effect of changing the metric.
A useful calibration report needs to distinguish three events: the renderer changed, the inputs changed, or the measurement changed. Only the first can directly support a claim that the app’s output improved. The other two can still teach us something valuable, provided we say what actually happened.
Related: why isolated slider improvements can fail in combinations, and how a photo editor can overfit to familiar photographs.
What this changes for Chiado
This change helps direct Chiado’s color work toward actual residual color and tone differences. Rechecks preserved the positive-blue selection improvement and the 17-of-18 native grading result, while identifying four changed classifications in the broader grading sweep. It is evidence that the acceptance process can correct its own measuring tool, rather than treating every lower score as a better image.