In practice, a photo can show skin tone without revealing skin type. Skin tone describes melanin amount and distribution, and how skin looks. Skin type describes function: sebum production, moisture retention, barrier integrity, and sensitivity. These are biologically independent dimensions and should remain separate in any recommendation system.
Many skincare recommendation systems still combine them, affecting who receives an accurate match. Separating them was an early Cutib design constraint, requiring deliberate decisions about model training and output structure.
Where the conflation comes from
Some confusion began with the way early clinical skin research was conducted. Much dermatology and cosmetic science literature drew on predominantly light-toned subject pools. Statements such as "oily skin tends to show X" or "dry skin presents as Y" were often based on a narrow visual baseline. Their descriptions therefore reflected, implicitly, how those traits appear on light skin.
Surface oiliness looks different on deeper skin tones, as does barrier disruption. Post-inflammatory hyperpigmentation, a common result of acne, sensitivity reactions, and product irritation, is more visible and consequential on deeper tones. PIH does not identify skin type. It indicates how skin responds to stress and inflammation. Yet systems that fail to separate these signals may send people with highly pigmented skin toward products selected from tonal features in a photo instead of functional ones.
What our model keeps separate
In a Cutib reading, skin tone and undertone are attributes related to formulation aesthetics: how a product appears on application, whether a tinted product suits the wearer, and whether certain color-active ingredients perform as expected. These inputs matter for some product categories.
Skin type attributes, including surface oil pattern, hydration levels, texture regularity, barrier markers, and sensitivity signals, come from separate image features. Our dataset deliberately covered a range of skin tones with independent variation in skin type. This taught the model that deep tone does not predict sebum production, and light tone does not predict dryness. The idea is simple, but implementation is harder: training data must contain that independence, while most publicly available skin datasets do not.
The post-inflammatory hyperpigmentation question
PIH merits specific attention because skin tone and recommendation logic do interact here. We had to define carefully what the model should infer, and what it should not.
PIH is not a skin type. It is a skin state indicator. In a photo, it may show a recent inflammatory event from acne, a product reaction, physical trauma, or something else. That helps guide recommendations: the barrier is likely recovering, current sensitization risk may be higher, and ingredient choices should favor anti-inflammatory and barrier-supportive options over actives that could further stress the skin.
PIH does not reveal baseline skin function. Someone with significant PIH may normally have non-sensitive skin after one reactive episode. They should not be permanently assigned to "sensitive skin" products because of their current complexion. Cutib treats PIH as a current-state signal, not a permanent classification. Recommendations should fit the present recovery state, not presumed long-term character.
The bias correction problem in training data
A key challenge in reading skin type across tones is that photographic texture and surface features vary with melanin level. Under the same lighting, identical sebum production can look different on light and deep skin. With unbalanced data, a model may learn these associations as the attributes themselves, rather than as tone-mediated appearances of tone-independent attributes.
We addressed this with targeted data collection and by making skin tone a conditioning variable for interpreting surface feature maps. Tone serves as context for functional signals, not as a predictor of them. This required more labeled data across the full tone range than planned, which is why early-access recruitment explicitly focused on demographic range.
This is not fully solved. We continue auditing accuracy by skin tone as more data arrives, and have found areas needing refinement. Equity in this domain requires ongoing work, not one training run. It was a design priority from the start, not a retrofit.
What this means for users and for brands
For users, Cutib should not infer skin function from tone. Deep skin tone does not mean oily, sensitive, or dry. Light skin tone does not mean fragile or easily dehydrated. Your profile comes from what your skin is doing in the photo, not from assumptions about what skin "like yours" usually does.
For brands, catalog coverage matters. If a range was formulated around a narrow tonal baseline, some products may not suit users outside it, even when skin type is compatible. Cutib does not currently assess formulation tone-appropriateness as makeup shade-matching systems do. It is a direction under consideration for the recommendation layer as catalog data becomes richer.
The principle is simple, though implementation is not: different skin tones can share a skin type, and one skin tone can include very different skin types. A system that ignores this independence will predictably fail many users in ways that can be corrected.