beauty API trained on Brazilian skin
Beauty API Trained on Brazilian Skin: Why the Dataset Changes the Result
Image analysis models inherit the diversity of the data they were trained on. Here's why a beauty API trained on Brazilian skin returns undertone, shade and hair attributes that are more useful for the catalog you sell here — and how to validate it before integrating.
When a brand evaluates a beauty API trained on Brazilian skin, the conversation usually starts with the endpoint, the JSON format and how long integration takes. Those are fair questions, but they come after a more important one: what kind of face was this model trained on? The answer decides whether the undertone field will help a foundation shopper complete the purchase or become one more data point the team learns to ignore.
Why the dataset matters more than the model's name
A photo analysis model learns to describe what it has seen. If most of the examples it saw sit inside a narrow band of shades, it gets good at that band and loses granularity outside it. In Brazilian e-commerce, that shows up in a very specific way: medium and deep shades get generic descriptions, olive and golden undertones get flattened into "neutral", and the shopper who most needs guidance is exactly the one who gets the least useful answer.
Brazil has a distribution of shades and undertones that is nobody's average. A foundation catalog here tends to carry more SKUs in the medium range than catalogs designed for other markets, and the difference between two neighboring SKUs is often the undertone, not the depth. An API that doesn't separate those cases well can't support a shade recommendation.
MaIA Scan was trained on Brazilian skin for exactly this reason. The output is a beauty profile in JSON: skin tone and undertone, skin type and visible condition, hair attributes and a confidence score per field. It describes visible attributes of a photo — it is not a medical assessment and does not replace a professional.
What actually changes in the API response
Undertone with commercial value
Undertone is the field that suffers most when the dataset lacks diversity. When it comes back well resolved, the product team can do something simple and valuable: map every foundation SKU in the catalog to a tone-and-undertone combination and let the API point to the right pair, instead of returning a list of ten options.
Visible condition described in your catalog's vocabulary
Apparent oiliness in the T-zone, a dry-looking appearance, visible marks: these are attributes the shopper recognizes in the mirror. When the reading is consistent across the whole customer base, CRM can build real segments and the product page can justify the recommendation with a sentence that makes sense.
Hair follows the same logic
Texture, curl pattern and strand appearance depend on data in the same way. A model with little variety of curly and coily hair tends to collapse everything into two or three categories — and a Brazilian hair care catalog does not fit into two or three categories.
The confidence score as a safety valve
No model gets everything right in every photo. That's why each field comes with a confidence level: the application decides whether to show the recommendation directly, ask for a new photo with better lighting, or offer an alternative path. That field is what separates an experience that fails silently from one that knows when to step back.
How to validate before integrating
Don't take the claim — not even ours — without testing. The most honest path is the simplest one:
- Build a photo set that looks like your shopper. Use images from your own audience, with the range of shades, ages, lighting conditions and camera qualities you actually receive. Studio photos don't represent the reality of mobile traffic.
- Define the ground truth first. Ask makeup artists or the product team to note the expected tone, undertone and attributes for each photo. Without ground truth, the evaluation becomes an impression.
- Compare by band, not by average. Split results by shade band. An API can look great on the overall average and fail exactly in the bands where your sales volume is concentrated.
- Look at behavior on bad photos. Yellow light, backlight, filters, heavy makeup. What matters isn't always being right: it's the confidence score dropping when it should drop.
You can run this test with purchased credits and an API key, with no long contract. The API documentation shows the request and response format; the selfie-based skin and hair analysis page details the profile fields.
From the profile to what the shopper sees
A well-resolved profile is an input, not a finished product. It becomes value when it feeds three fronts:
- Shade recommendation on the product page, in site search and in the repurchase email, pointing to SKUs from your catalog.
- Visual preview, with virtual makeup try-on: in the API format, the photo comes back with the lipstick or foundation rendered in your palette's real colors; in the browser SDK, in early access since September 2026, the shopper tries thirteen products and the Full Look in live camera mode, with processing happening locally, or in photo mode.
- Customer service, with MaIA Beauty Advisor reading the profile and talking to the shopper about products from your catalog on WhatsApp, on the page or in chat — delivered as a project with the B4A team.
What to ask any vendor
Whatever vendor you're evaluating, three questions resolve most of the doubt: what population the model was trained on; which fields come with a confidence indicator; and what happens to the photo after the analysis. If the first answer is vague, the other two matter less — the result will already arrive biased at the end of the line.
Want to test it with your own photos? Create an account, buy credits and generate your API key, or talk to the B4A team to design a pilot around your catalog.
Frequently asked questions
Why does it matter that the API is trained on Brazilian skin?
Photo analysis models describe well what they saw often during training. With data concentrated in a narrow band of shades, olive and golden undertones and medium and deep tones get generic descriptions — exactly where the Brazilian foundation catalog has the most SKUs.
How do I test quality before integrating?
Build a photo set from your own audience, define the expected tone and undertone with your product team or makeup artists, and compare results by shade band rather than by overall average. Also run photos with poor lighting to see whether the confidence score drops when it should.
What comes back in the MaIA Scan response?
A JSON with the beauty profile: skin tone and undertone, skin type and visible condition, hair attributes and a confidence score per field. These are visible attributes of the photo, not a medical assessment, and they do not replace a professional.
Can I use the profile together with virtual makeup try-on?
Yes. The profile indicates the tone and undertone range and the try-on shows the result: in the API format, the photo comes back with the product rendered in your palette's real colors; in the browser SDK, in early access, the shopper tries it in live camera mode or in photo mode.
Do I need a contract to start testing?
No. MaIA Scan and MaIA Try-On via API are self-serve: you create the account, buy credits and generate the API key. MaIA Beauty Advisor is delivered as a project with the B4A team.