Color-ai-models

Color AI — chroma classifiers

Three trained color-naming classifiers: given a color, they name it — "red", "navy", "charcoal", etc. Each is a scikit-learn RandomForestClassifier, trained entirely on synthetic labeled data (no photos, no dataset needed to build them) — see color-ai-core on PyPI for the code that trained these and the functions to use them correctly.

Files

Nano covers the basics: red, orange, yellow, green, teal, blue, purple, pink, brown, black, gray, white.

Micro and Micro Super add 12 more: maroon, navy, olive, lime, cyan, magenta, indigo, coral, mint, lavender, beige, charcoal. Micro Super is the same 24 classes as Micro, just trained on more samples per class (800 vs 500), which measurably improves accuracy — same anchors, same architecture, purely more training evidence.

Micro and Micro Super also support a two-stage fallback: when the model isn't confident (below 50%) in its specific pick, it's meant to report the broader Nano-family color instead of guessing between two close options (e.g. "brown" instead of a shaky call between beige and charcoal). This logic lives in the color-ai-core package, not baked into the .joblib file itself — see Usage below.

Usage

⚠️ These models do not take raw RGB as input. They were trained on a 4-value feature vector — [sin(hue), cos(hue), saturation, value] — not [R, G, B] directly. Feeding raw RGB straight to .predict() will run without error but give meaningless results. Use the matching feature function from color-ai-core rather than hand-rolling it:

pip install color-ai-core huggingface_hub

from huggingface_hub import hf_hub_download
from color_ai_core import load_model, predict_color

path = hf_hub_download(repo_id="YOUR_USERNAME/YOUR_REPO_NAME",
                        filename="color_classifier_micro.joblib")
clf = load_model(path)

name, confidence, used_fallback = predict_color(clf, [220, 40, 30], tier="micro")
print(name)  # "red"

predict_color() handles the RGB → feature conversion and the two-stage fallback correctly — that's why it's the supported way to use these files rather than calling clf.predict() directly.

Training

All three were trained on jittered synthetic data generated around fixed HSV anchor points per color (e.g. red ≈ hue 0°, high saturation, high value) — noise is added to hue, saturation, value, and the final RGB to simulate lighting and camera variation, so each model learns a color's range rather than one exact shade. No photos were used to train these specific files. color-ai-core also includes train_from_images() for training a classifier on your own labeled photos instead, if you'd rather have that.

License

MIT