title: Color for ML: Task and Evidence Contract description: "Color guidance for ML is a task-specific representation and evidence contract: name the source encoding, illuminant or viewing assumptions, target transform, palette intent, and human-review purpose; artistic harmony, spectral labels, and psychological associations are hypotheses, not universal labels or model controls." category: reference tags: [color, machine-learning, colorimetry, palette, data-provenance, evaluation, color-management] aliases: ["Color Theory", "Color Reference for Diffusion"]
Color for ML: Task and Evidence Contract¶
Color in an ML workflow is not one universal feature. A palette description, device RGB values, a calibrated colorimetric measurement, and an editorial preference answer different questions. Name the representation and the target task before turning color into a training signal, control, score, or quality claim.
Separate the evidence types¶
| Use | What it can support | What it cannot establish alone |
|---|---|---|
| Color-managed capture or target | a measured transform under declared conditions | a universal appearance under every illuminant or display |
| Palette or harmony brief | approved editorial intent | an objective quality or psychological truth label |
| Prompt or reference image | a requested visual direction | deterministic color control across model releases |
| Annotator preference | a bounded review decision with rubric and context | a universal human response or protected-trait inference |
CIE Colorimetry describes standardized colorimetric foundations and references color-appearance models. That supports explicit representation and viewing assumptions; it does not make artistic wheel positions, wavelength names, or cultural associations universal labels for a model.
Bind the data representation¶
For every training or evaluation set, record:
- source file/profile, decode, transfer function, and working representation;
- illumination or viewing assumptions when measured appearance matters;
- any ICC, OpenColorIO, or other transform/configuration revision;
- crop, resize, tone mapping, gamut mapping, and color augmentation policy;
- palette/reference provenance, authorial intent, and rights; and
- label rubric, annotator context, uncertainty, and exclusion policy.
The ICC profile architecture and OpenColorIO configuration model make transforms explicit. Apply the same discipline to data ingestion: unlabeled RGB values from mixed sources must not be treated as one calibrated color domain.
Design controls for the named task¶
A palette controller can be evaluated against an approved target palette or reference image. A color-correction model can be evaluated against source-aligned measurements. A generative model can be reviewed for whether its output follows a permitted visual direction. Each needs a different acceptance test.
Do not assume that a color name maps to a fixed device value, that a text prompt controls a hue identically across checkpoints, or that a wide-gamut intent survives an sRGB-only source/output path. Pin the model, tokenizer or prompt processor, image pipeline, and output transform when making a reproducible color claim.
Evaluate without inventing human facts¶
Split data by original asset, capture session, derivative chain, and reference palette so near-duplicates cannot leak color targets into holdouts. Test on the actual delivery transform and review gradients, neutral/target regions, gamut behavior, material identity, text, and protected areas.
Color may influence an editorial review, but do not use it to infer health, emotion, ethnicity, attractiveness, or another sensitive trait. Psychological or cultural color associations are contextual hypotheses that require their own consent, study design, and bounded interpretation; they are not automatic training labels.
Release gate¶
Release a color-aware ML capability only with an identified representation, traceable transforms, authorized data, source-disjoint evaluation, output provenance, and task-specific human review. If the workflow cannot show whether a result is a calibrated correction, a requested style, or plausible generation, label it as a reviewable visual proposal rather than factual color recovery.