On this page

Dominant Colour

Colour is one of the first things anyone notices about a picture and one of the last things most media systems can actually search on. This node closes that gap: it works out which colours a photo is really made of, and reports each one both as a precise value and as a word a person would use.

So the same result gives you #E2711D for your design tools, and "vivid orange" — or kräftiges Orange — for your search box, your caption prompt and your editors.

It runs no model. All of it is arithmetic, so it costs no GPU time, needs nothing installed alongside it, and finishes in milliseconds.

Kind

dominant-color

Applies to

Images

Inputs

The picture. Optionally bounding boxes from Face Detection or any other detector

Output keys

dominant_color_result (JSON String), dominant_color_hex, dominant_color_term, dominant_color_name_en, dominant_color_name_de (Text), dominant_color_region_count (Number)

Requirements

None — no model, no sidecar, no GPU

Persists to

asset_json_comp — one row per asset, holding every region

What it produces

For each area it measures: the dominant colour, the ranked palette behind it, and how much of the area each colour accounts for.

{
  "image": { "width": 1920, "height": 1080 },
  "regions": [
    {
      "id": "whole", "kind": "IMAGE", "pixels": 39204,
      "dominant": {
        "share": 0.4712,
        "hex": "#E2711D",
        "rgb": { "r": 226, "g": 113, "b": 29 },
        "hsl": { "h": 30.5, "s": 77.3, "l": 50.0 },
        "lab": { "l": 60.13, "a": 38.21, "b": 62.35 },
        "lch": { "c": 73.13, "h": 58.5 },
        "name": { "term": "orange", "en": "vivid orange", "de": "kräftiges Orange" }
      },
      "palette": [
        { "share": 0.4712, "hex": "#E2711D", "name": { "en": "vivid orange", "de": "kräftiges Orange" } },
        { "share": 0.3105, "hex": "#6B4E8C", "name": { "en": "dark purple", "de": "dunkles Violett" } },
        { "share": 0.2183, "hex": "#F2C57C", "name": { "en": "light yellow", "de": "helles Gelb" } }
      ]
    }
  ]
}

The share values tell you how much of the picture each colour covers, so you can render a real proportional swatch strip rather than a list of equal chips.

Three ways to give a colour a name

Each colour comes back with three labels, and they are meant for different jobs:

Field Use it for

term

Filtering, facets and grouping. One of eleven stable words — red, orange, yellow, green, blue, purple, pink, brown, black, grey, white — that never changes when the wording is refined

en / de

Showing to people, and feeding caption prompts. Fuller descriptions like "dark greyish blue" or dunkles graustichiges Blau

hex / rgb / hsl / lab

Design tools, palettes and exact comparisons

Why the names sound like a person picked them

Computers usually name colours by finding the closest entry in a long list, which produces "lightsteelblue3" and similar. This node does something closer to what people do: it decides the basic colour first, then describes it.

The basic colour is chosen in a perceptual colour space — one built so that the distance between two colours matches how different they look to a human eye, which is not true of raw RGB. The description is then built from how light and how saturated the colour is, giving pairs like "pale blue" and "deep blue", or blasses Blau and tiefes Blau.

Two consequences worth knowing:

  • Greys, blacks and whites are never given a hue. Below a saturation threshold the node stops asking "which colour is this?" and answers on brightness alone, so you get "light grey" rather than a technically-correct-but-absurd "light greyish orange".

  • German is generated, not translated. Both languages come from the same decision, so they can never drift apart or describe the same colour differently.

Colouring in parts of a picture

Point a detector at the image first and the node measures every box it finds, alongside the whole frame. That turns "this photo is mostly blue" into "this person is wearing blue".

You can also pin it to a fixed area — a logo corner, a lower third, a product region — without any detector at all, using the region options below.

Configuration

DominantColorNodeOptions:

Option Meaning

clusterCount

How many colours to pull out of each area. Default 5

maxSamples

How many pixels to look at per area. Default 40,000 — enough for a stable answer on any size of picture

includeWholeImage

Measure the whole frame. On by default

useDetections

Measure every detected box as well. On by default

detectionSources

Which upstream detector nodes supply the boxes. Defaults to facedetect

regionX / regionY / regionW / regionH

A fixed area to measure. Given as fractions of the picture by default; set regionCoordinates to ABSOLUTE_PIXELS to use pixels instead. Leave the width and height at 0 to switch it off

alphaThreshold

How opaque a pixel must be to count. Transparent areas are ignored rather than treated as white, so a logo on a transparent background reports its own colours. Default 128

minRegionPixels

Areas smaller than this are skipped — there is nothing meaningful to measure. Default 64

maxRegions

Upper bound on detected areas, largest first. Default 32. Anything dropped is reported in the result

emitPalette

Return the full ranked palette, not only the top colour. On by default

achromaticChroma

How unsaturated a colour must be before it is called grey, black or white instead of being given a hue. Default 12

blackLightness / whiteLightness

Where black ends and white begins on the brightness scale. Defaults 20 and 85

seed

The same picture always produces the same palette. Change this only if you want a different-but-still-repeatable result

Use Cases

  • Search and filter by colour — "show me the orange product shots", "find the photos with a green background".

  • Brand compliance — check that assets actually use the brand palette, and flag the ones that drift.

  • Better captions — feed "vivid orange sky, dark purple hills" into a caption prompt instead of hoping the model notices.

  • Automatic theming — pull a page or player accent colour straight from the hero image.

  • Design handoff — hand editors a real palette with proportions, in HEX and in words.

  • Bilingual catalogues — the same asset is searchable as "blue" and as Blau with no extra work.

Good to know

  • Video is not supported yet. Colour changes shot to shot, so a single answer for a whole clip would be misleading. Run it on stills, or on extracted frames.

  • A picture that cannot be opened is reported as a failure, not quietly skipped — so a corrupt file shows up in the run summary instead of disappearing. CMYK JPEGs are the common case Java cannot decode.

  • Cyan and teal are reported as blue, and turquoise as light blue. The eleven basic colour words are chosen because they exist in every language; finer shades are always available as HEX.

  • The result is repeatable. The same picture and settings always give the same palette, so cached results and search facets stay stable.