Every tool that pulls a colour palette out of an image is making a claim about perception, usually without meaning to. Cluster the pixels in RGB, take the cluster centres, call them the dominant colours. The output is defensible arithmetic and frequently looks nothing like the palette a person would pick.
This paper attacks the problem at the only place it can be fixed: where the clustering starts, and how distance is measured.
It was posted to arXiv on 23 October 2026 (arXiv:2610.32473) by Uldana Shyndali, Nurdaulet Toleugaliyev, Kuat Zharykbayev, Zarina Abildasheva and Pakizar Shamoi. The Comments field is empty and there is no journal reference, so this is an unreviewed preprint.
The diagnosis
The paper’s statement of the problem is worth having in front of you, because it names two separate faults:
conventional algorithms based on device-oriented color spaces (e.g., RGB) and fixed quantization often fail to capture how humans perceive color differences
Device-oriented colour space. RGB describes what a display emits, not what an eye resolves. Equal steps in RGB are not equal steps in appearance — the green channel carries far more perceived lightness than the blue, so Euclidean distance in RGB systematically mis-ranks which colours are “close”.
Fixed quantization. Chopping the colour cube into a regular grid puts boundaries where the geometry is convenient rather than where perception has joints.
Both faults push in the same direction: toward palettes that are statistically correct and visually unconvincing.
The experiment
Rather than derive perceptual anchors from a model, they asked people.
110 participants were shown 120 colours, sequentially, and asked to identify those they perceived as unique. From the responses the authors built a 120 × 120 co-occurrence matrix — which colours were jointly judged distinct — and applied a greedy selection algorithm to extract 64 perceptually distinct colours. That set is the Human Color Dictionary (HCD).
The design is neat because the question asked of participants is not “name this colour” or “are these two the same”, both of which import vocabulary and task effects. “Which of these look unique to you” produces a judgement about distinctness directly, and the co-occurrence structure does the work of turning 110 individual judgements into a set.
How it is used
The HCD is dropped into two standard algorithms rather than replacing them:
- RGB colour histogram extraction
- K-means extraction
In both, the dictionary is used to initialise the centroids, and pixels are mapped to their nearest perceptual colour using ΔE₀₀ — the CIEDE2000 difference formula — instead of Euclidean RGB distance.
That is the whole intervention: better starting points, and a distance metric built for appearance rather than for storage. No new architecture, no training.
The evaluation
Against 4,562 images paired with expert-designed palettes, scored with Earth Mover’s Distance, the method “improves perceptual alignment”.
The choice of EMD is the right one and worth a note. Comparing two palettes is not a matter of matching colours one-to-one — a good palette may use four colours where the reference used five, and be no worse. Earth Mover’s Distance measures the cost of transporting one distribution of colours into another, so it rewards being close overall rather than being identical item by item. It is the metric that corresponds to “does this palette feel like that palette”.
Having expert-designed palettes as ground truth is the real asset here. Most palette-extraction papers evaluate against other algorithms, which only establishes that two pieces of arithmetic disagree.
What to be careful about
No effect sizes in the abstract. “Improves perceptual alignment” is a direction, not a magnitude, and with an empty Comments field and no venue there is no reviewer standing behind it either. Read the paper before quoting a number, because there is no number here to quote.
64 colours is a design decision, not a discovery. The greedy selection produced 64 from 120 candidates; a different stopping point would produce a different dictionary. That the number lands on a power of two is convenient rather than meaningful.
110 participants, 120 colours, one sitting. Colour judgement is affected by display calibration, ambient light, viewing order and, at the population level, by language and culture — the vocabulary available for colour categories differs between languages in ways that are well documented. A dictionary derived from one participant pool is a reasonable starting point and not a universal perceptual constant. The paper does not claim otherwise.
Why this is useful if you make pictures
Three concrete takeaways that do not depend on the paper’s numbers holding up.
If you are writing palette extraction, your distance metric is a bug. Swapping Euclidean RGB for ΔE₀₀ in CIELAB is a contained change, the formula is published, and implementations exist in every language. It will change your output on exactly the cases you currently get wrong — near-neutrals, saturated blues, and anything where two colours are numerically close and visually distant.
Seeding matters more than iterating. K-means converges to a local optimum determined by where it starts. Perceptually chosen seeds are a cheaper intervention than a better clustering algorithm.
“Dominant” is ambiguous, and worth resolving on purpose. The most frequent colour, the most perceptually salient colour and the colour a designer would name as the image’s identity are three different things. An automatic extractor returns the first. This work moves the output toward the third. Which one your tool should return depends on whether it is building a thumbnail background or suggesting a brand palette.
Sources
- From Perception to Pixels: Human-Derived Visually Distinct Colors for Improved Dominant Color Extraction — arXiv:2610.32473
- arXiv PDF
- CIEDE2000 colour-difference formula overview — Bruce Lindbloom
- Earth Mover’s Distance