wordpalette

A ~710k-parameter character-level model that predicts the colours a word evokes. You give it a word; it returns up to three CIELab colours with proportions, and decides for itself how many a word has (blood → one red, france → a blue/white/red tricolour). Trained entirely on free, human-made colour data — no large language model was used at any point, for labels or distillation.

• Demo: https://alganet.github.io/wordpalette/
• Code & training pipeline: https://github.com/alganet/wordpalette
• Size: 2.8 MB ONNX (fp32), runs on CPU / in-browser via onnxruntime-web
• License: CC-BY-4.0 (weights) · code is MIT

What it's for

Generative/affective colour: pick a theme colour from a tag, colour a word cloud, seed a palette from a brand or a mood, or just explore how spelling maps to colour. Because the encoder is character-level, it colours words it has never seen — arborious → green, crimsonish → red — from morphology alone.

Usage

Inference needs only numpy + onnxruntime (no PyTorch):

import wordpalette                      # pip install wordpalette
wp = wordpalette.load()
[(s.hex, round(s.proportion, 2)) for s in wp.palette("brazil")]
# [('#30a432', 0.69), ('#ffd641', 0.25)]

Or drive the ONNX file directly:

import json, numpy as np, onnxruntime as ort
vocab = json.load(open("vocab.json"))
sess = ort.InferenceSession("wordpalette.onnx")

def encode(w):
    ids = np.zeros(vocab["max_len"], dtype=np.int64)          # PAD = 0
    for i, c in enumerate(w.lower()[:vocab["max_len"]]):
        j = vocab["chars"].find(c)
        ids[i] = (j + 2) if j >= 0 else vocab["unk"]
    return ids

lab_scaled, props = sess.run(None, {"chars": encode("ocean")[None]})
lab = lab_scaled[0] * np.array(vocab["scale"])                # [K, 3] real CIELab

I/O. Input chars: int64 [batch, 24], characters mapped as a–z + #→2.., others→UNK(1), zero-padded (PAD 0). Outputs lab_scaled [batch, 3, 3] (multiply by scale = [100, 110, 110] for real Lab) and props [batch, 3] (softmax proportions). A slot is "on" when its proportion ≥ 0.15; always keep at least the dominant slot.

How it was trained

Ten free colour sources — Simple Icons, Company Colors, color-names (meodai), XKCD, CSS, GitHub label colours, emoji dominant-colour, flag palettes, WordNet glosses, and Wikipedia first sentences — are normalised to word → CIELab with a trust weight, then coalesced into one belief per word (aggregate GitHub duplicates, denoise over-segmented palettes, drop cross-source outliers, demote logo colours that contradict a word's everyday sense). A causal GRU is then trained per-prefix in scaled Lab with:

• a hybrid loss — crisp where multi-colour ground truth exists (flags), flexible where it doesn't (single-colour words keep plausible secondaries);
• card-trust — WordNet teaches hue but not colour count;
• EMA weight-averaging for seed-robustness.

The full, runnable pipeline (and the reasoning behind each choice) is in the GitHub repo. Reproduction is a short CPU/GPU job from the committed data.

Evaluation & limitations

Evaluated during development against the NRC/Mohammad word–colour association lexicon (held out, never trained on) plus flag-palette accuracy and behavioural probes. It is a small model over noisy, partial supervision: it captures volume and patterns (thousands of natural kinds, morphology, flags) but not isolated facts (a handful of planets or trademark-only brands may be approximate). Colours are associative and sometimes contestable — treat outputs as evocative, not authoritative.

Attribution

Weights derive from the sources listed above; please credit them (see the repo README for per-source licences). Notably the WordNet License (Princeton) and CC-BY-SA text from Wikipedia contribute; no source trained an LLM.

Citation

@software{wordpalette,
  author = {Gomes Gaigalas, Alexandre},
  title  = {wordpalette: a tiny character-level word-to-colour model},
  year   = {2026},
  url    = {https://github.com/alganet/wordpalette}
}