We trained a compact AI model on 921 multispectral images captured with our Lumoscope. It reads age from skin as accurately as two of the largest AI systems in the world at a tiny fraction of their size. It pairs a standard, off-the-shelf image model with a simple formula trained on our scans, and it runs locally at essentially zero cost per scan, with no API fees.
Shown scans from two people at least 20 years apart, our model picked the older person 78% of the time, and 71% of the time when they were at least 10 years apart. On the 20-year test, Anthropic’s Claude scored 71% and OpenAI’s GPT-6 Astra 72%. Estimating exact age, our model missed by an average of 6.34 years, ahead of Claude at 6.42 and GPT-6 Astra at 6.66. The margins between the three fall within statistical noise. Still, a model small enough to run on a laptop, trained only on our own scans, matched or edged out general-purpose systems trained on vastly more data on every measure.
The study drew on 921 multispectral images (308 scans) from 289 consenting adults aged 18 to 54, each scanned and skin-typed in person by a trained operator. Every estimate was checked against the person’s recorded age. A second Lumeria researcher then rebuilt the experiment from scratch on a separate computer and matched every headline number to two decimal places.
| Model | 20+ years apart | 10+ years apart | Ranking score |
|---|---|---|---|
| Lumeria’s model | 78% | 71% | 0.63 |
| Claude (Anthropic) | 71% | 67% | 0.62 |
| GPT-6 Astra (OpenAI) | 72% | 66% | 0.60 |
| Coin flip | 50% | 50% | 0.50 |
What the numbers mean
Average age difference is the simplest way to grade an age estimate. It’s the average of all differences between true age and estimated age. We ran the test for each person’s skin and then averaged the differences, and you get the score in the table. Most models are able to achieve less than 7 years of average difference.
Guessing the median age is the baseline every model has to beat. Line everyone in the study up from youngest to oldest and take the person in the middle. If you ignored the scans completely and guessed that one age for everybody, you would be off by 6.87 years on average. A model that can’t beat that number hasn’t learned anything from the skin. All three models beat it, and ours cleared it by the widest margin.
| What made the estimate | Average difference |
|---|---|
| Lumeria’s model | 6.34 years |
| Claude (Anthropic) | 6.42 years |
| GPT-6 Astra (OpenAI) | 6.66 years |
| Baseline: guess age 29 for everyone | 6.87 years |
These numbers can look larger than some published skin-age results, and the reason is the age range. Our participants span 18 to 54, where a single middle guess is already fairly close for most people. Studies covering ages 20 to 80, such as a 2024 study of hand skin that reported about 4.7 years of difference (Georgievskaya et al., Experimental Dermatology), are grading a much wider spread. “Difference” figures only mean something next to their own baseline, so we always publish ours alongside the result.
What is an age estimate you might ask? If a model guesses 40 for someone who is 35, the person’s skin looks younger by five years. Part of every age estimate is a person’s skin being older or younger than their years, and that gap is exactly what a Lumeria scan is built to show you and help you tackle. We plan to use this to give you an understanding of where your skin falls relative to people your age, and how that changes as you adjust your routine.
We ran four checks to confirm the gain comes from the skin itself. An untrained copy of the same network, with random weights, scored right at the baseline, showing the improvement comes from what the model learned rather than its architecture. In a permutation test, we randomly reassigned ages across scans and retrained 200 times, and none of the shuffled runs matched the real model (p = 0.005). The result also held on collection days and scan operators the model had never seen, and on black-and-white versions of the scans, ruling out skin color as the signal.
How the test worked
Each Lumoscope scan is three images of the same patch of skin under controlled light: ordinary white light, cross-polarized light (which filters out surface shine to show what sits just beneath it), and long-wave ultraviolet light. Because the lighting, distance and surrounding environment are fixed for every scan, differences between scans come from the skin rather than the room.
No one in the test set appeared in the training set, so every score reflects how the model performs on new users rather than on people it has already seen.
Claude and GPT-6 Astra received the same 209 scans at full resolution, were told the age range of the group, and were asked three times per scan to estimate age from the skin alone. All three models then went through the same standard calibration step, which corrects for a system that consistently guesses a few years high or low. Every one of these rules was set before any results came in. Running Claude and GPT-6 Astra on the test cost about $30 in API fees; our model scored the same scans locally for free.
We also tested a widely used open-source method that scores skin with traditional image analysis, counting edges for wrinkles and measuring color for redness. It works on ordinary face photos, but at the Lumoscope’s close-up magnification its wrinkle detector mostly traced individual hairs, and its scores had essentially no link to age. Methods built for one kind of camera have to be re-tested on a new instrument before anyone should trust them.
What the model is picking up on
Because the model held up on black-and-white scans, across skin tones, and on operators and days it had never seen, what it reads is the texture and structure of the skin itself, not its color, the lighting, or who ran the scan.
The light mattered too. Hiding the fine texture in the cross-polarized image moved the model’s estimates about twice as much as doing the same to the white-light or UV images. The model relies most on the cross-polarized image: the view that filters out surface shine to reveal the texture just beneath it, something an ordinary phone camera can’t capture.
Our hand-built measurements hint at which structures matter. Of nine classic methods we tried for picking out fine lines, two tracked age best: a ridge filter that finds the shallow furrows in the skin’s surface, and a measure of how aligned the skin’s micro-texture is. Young skin has a fine, criss-crossing texture that lines up into parallel furrows with age, one of the most consistently replicated findings in skin-aging research. These results are early, and they describe our hand-built measurements rather than the neural model directly.
What comes next
The first benchmark also shows where more data will help most. The model is most precise where our dataset is deepest: for people aged 25 to 34, the largest group in the study, it is off by about 3 years on average. Only 25 participants so far are aged 45 to 54, so for now the model tends to underestimate age for that group, placing its estimates closer to the middle of the range. Expanding our older cohort is the most direct way to sharpen those readings.
Skin tone is the other gap. Every participant’s skin type was assessed by a trained operator at the time of scanning, using the Fitzpatrick scale, the standard six-point dermatology scale that groups skin by tone and by how it responds to sun. Of the 290 people scanned so far, 91% fall in the middle three types, II to IV. Only 7 people are type I, the fairest skin, and just 14 are types V and VI, the deepest.
| Fitzpatrick type | People | Share | Test scans |
|---|---|---|---|
| I: very fair, always burns | 7 | 2.4% | 5 |
| II: fair, usually burns | 83 | 28.6% | 61 |
| III: medium, sometimes burns | 130 | 44.8% | 101 |
| IV: olive to light brown, rarely burns | 52 | 17.9% | 35 |
| V: brown, very rarely burns | 10 | 3.4% | 4 |
| VI: deep brown to black, never burns | 4 | 1.4% | 2 |
| Not recorded | 4 | 1.4% | 1 |
With only a handful of scans at either end of the scale, we can’t yet say how accurate the model is for the lightest and darkest skin, and we won’t make that claim until we can back it up. We’ve noticed this skew, and we are looking for volunteers with lighter or darker skin to diversify our dataset and make sure the model is accurate for everyone.
Our next round of data collection targets both gaps directly. We are recruiting more participants over 45, including people over 55 who fall outside today’s study, and more participants at both ends of the Fitzpatrick scale. Every new scan is captured the same controlled way, so the model gets better with each one, and we will publish the updated benchmark, including accuracy for each skin type, in the same format when it is ready.
Questions?
If you have questions about the benchmark, the methods, or volunteering for our next study, email us at contact@lumeria.skin. We read every message, and we especially want to hear from people with very fair or deep skin, and from anyone over 45.
Is a skin-age estimate a diagnosis?
No. This is a research benchmark of Lumeria’s imaging and models, and a skin-age reading is not a medical measurement. Lumeria offers insight and monitoring, not diagnosis. For any concern about your skin, see a licensed dermatologist.