How Lumière reads every skin tone

The bias problem (in one paragraph)

A series of audits since 2018 have shown that consumer and clinical dermatology AI tends to perform worse on darker skin. The cause is unsexy: training datasets that over-represent lighter skin (Fitzpatrick I–III) and under-represent deeper tones (Fitzpatrick IV–VI). Models inherit the dataset. The same model that confidently identifies erythema on Fitzpatrick II skin can miss post-inflammatory hyperpigmentation on Fitzpatrick V skin — or label it as something else entirely. This is not a conspiracy. It's a data problem with a downstream consequence: when a Black woman opens a skincare app and gets "you have oily skin" as her entire analysis, the app isn't being lazy. It's being honest about what it can see.

Lumière can't fix the upstream bias problem in dermatology AI as a whole — that's a category-wide data issue that takes years and tens of millions of dollars of ethically-collected ground truth to undo. What we can do is build our analysis to commit to a tone classification first, every time, and interpret findings in that context rather than against a default assumption. That's what this page documents.

What "calibration" actually means here

Lumière's skin analysis is anchored to peer-reviewed clinical frameworks — Fitzpatrick phototyping, the Monk Skin Tone scale, GAGS for acne, mMASI for melasma, ITA colorimetry. The analysis is required to commit to a tone classification before it interprets pigmentation, vascular signs, or barrier health. That ordering is the difference between a finding read in context (a brown spot on Fitzpatrick V skin is most likely PIH) and one read against a default (a brown spot read as "hyperpigmentation, see a dermatologist").

Tone-first interpretation is not what most consumer skincare AI does. It's the single biggest reason a generic "skin analyzer" can return a useful read on Fitzpatrick II skin and a flat one on Fitzpatrick V.

The calibration anchors, in detail

1. Fitzpatrick I–VI typing

Every scan returns a Fitzpatrick phototype (I through VI), based on visible undertone, melanin density, and an estimate of how the skin would respond to UV exposure. This is the coarsest classification we use, but it's the one most clinicians recognise — so it's the one our analysis is anchored to first.

2. Monk Skin Tone scale (1–10)

Fitzpatrick has six rungs. The Monk Skin Tone scale, which Google open-sourced in 2022, has ten — and it's a much finer-grained classification that captures the deeper end of the spectrum more honestly. Lumière requires every scan to land on a Monk tone (1–3 = very light, 4–6 = mid-tone, 7–10 = deep / rich). This second, finer classification is what lets the analysis distinguish between, for example, Monk 7 and Monk 9 — two tones that Fitzpatrick V flattens into a single label.

3. ITA (Individual Typology Angle) + overtone + undertone

ITA is a colorimetric measure used in cosmetic science to place skin on a continuous spectrum from "very light" to "dark." We surface an ITA estimate alongside Fitzpatrick and Monk so the analysis can ground its findings in a continuous measurement, not just categorical buckets. Overtone and undertone are surfaced separately — recognising that two people at the same Monk tone can have very different undertones (cool, warm, neutral, olive), which changes what tinted SPFs and concealers will flatter them.

4. Pigmentation: PIH vs PIE vs melasma vs lentigines

This is the most important calibration we do, and the one most consumer skincare apps get wrong on darker skin.

• PIH — post-inflammatory hyperpigmentation. Brown / dark spots left after acne or injury. Far more common on Fitzpatrick IV–VI.
• PIE — post-inflammatory erythema. Red / pink marks left after acne or injury. Far more common on Fitzpatrick I–III.
• Melasma — symmetric brown patches, usually on the face, often hormonal in origin.
• Solar lentigines — sun-driven dark spots, distinct from PIH.
• Periorbital pigmentation — under-eye darkness, which is often vascular, structural, or pigmented; treating these the same way is why most "dark circle" creams don't work.

Most consumer skincare AI returns "hyperpigmentation" as a single finding. Lumière differentiates between the categories above on every scan. That distinction matters because the treatments diverge: PIH responds to brightening agents and sun protection; PIE responds to vascular-targeted ingredients (azelaic acid, niacinamide, gentle laser); melasma responds to a third protocol entirely.

5. Vascular signs across skin tones

Rosacea, telangiectasia, and chronic flushing are consistently under-diagnosed on darker skin — partly because "redness" reads differently against deeper melanin (often as warmth, swelling, or a violaceous tint rather than visible erythema). Lumière looks for vascular signs through tone-appropriate cues, not only through the visible-redness shortcut. This alone is the difference between "no findings" and "your skin is showing early rosacea presentation" for many Fitzpatrick IV–VI users.

6. Acne grading (GAGS) + lesion typing

Acne presents differently across skin tones — not in severity, but in what it leaves behind. PIH after a single breakout on deep skin can outlast the breakout itself by months. Our acne profile uses the Global Acne Grading System (GAGS) and counts active lesions separately from residual marks, so a user with a clear-but-marked face isn't flagged as having severe active acne.

7. Barrier health: dehydrated vs dry

Dehydration (water deficit, transient, presents as tightness and dullness) and dryness (lipid deficit, chronic, presents as flakiness and rough texture) are mechanistically different and respond to different ingredients. Lumière commits to one or the other on every scan, rather than lumping both into "dry skin."

What we won't claim

A methodology page should be as honest about its limits as its capabilities. Lumière is not:

• A medical device. Lumière's analysis is informational, not diagnostic. We surface estimates of visible skin signs. We do not diagnose disease, we don't prescribe, and we will tell you to see a dermatologist when something we flag warrants it.
• Perfect at every camera, every angle, every lighting condition. Bad lighting — especially warm tungsten or harsh blue LEDs — skews tone classification. Lumière surfaces a confidence score on every scan and asks you to retake when confidence is low. If a scan looks wrong, it probably had a low confidence number we should have surfaced more loudly — and that's on us, not on you or your skin.
• Replacing your dermatologist. If you have a persistent concern — a mole that's changing, a rash that won't resolve, severe acne, anything that hurts — see a clinician. Lumière is a daily companion, not a substitute for medical care.

Five things you can hold us to

1. Every scan returns a Fitzpatrick phototype AND a Monk Skin Tone classification. Not optional. Not "skin tone unclear." If you ever see a scan output that doesn't include both, that's a bug — email us with the date and we'll investigate the specific scan.
2. PIH and PIE are surfaced as separate findings, not blended into "hyperpigmentation." If your scan returns "hyperpigmentation" as a single bullet without specifying which kind, that's also a bug.
3. Vascular detection runs on every scan, regardless of skin tone. Lumière looks for rosacea and telangiectasia signs through tone-aware cues, not only visible redness.
4. We publish our calibration anchors here. This page. If we change them — add new ones, deprecate old ones, refine how the system reasons — we update this page, dated.
5. If your scan reads your skin wrong, we want to know specifically what it got wrong. Not a five-star rating, not a one-star rating — the scan, the specific finding, and what the right answer would have been. We use that signal to refine our analysis and to flag scans for closer review.

If your scan didn't read your skin right

Email support@lumiere-skin.us with subject line "Methodology feedback" — include the date of the scan, what it returned, and what it missed. A photo of the scan output is helpful but not required. The founder reads every one — this isn't a ticket queue.

We don't get this right every time. We do try to get it less wrong, every release.

Versioning

This methodology was last updated alongside Lumière app version 1.1.2 / Build 42. Substantive changes to the calibration anchors — new tone scales, new pigmentation categories, new vascular subtypes — will be listed here and dated. Cosmetic changes to this page will not.