PinkMirror review: how its beauty score is built, what we could and could not verify, and how to check any face rating

PinkMirror is one of the most searched names in AI beauty scoring, and almost nothing written about it separates what the service says about itself from what other sites claim about it. That matters, because the most-copied claims online are about its price and its 'ratings', and those are the two things nobody checks. This review does the opposite: it uses PinkMirror's own pages for what the product is and how the score is built, says plainly which pages we could not read, and then shows how to test any face score, PinkMirror's or ours, for accuracy.

Search results for the brand are PinkMirror's own pages, its social profiles and a review-site listing. There is no independent write-up, so the first useful thing to do is separate what the service says from what is repeated about it. The table lists every claim in this article that comes from PinkMirror, and where it comes from.

Two things are worth noticing. First, the product is a browser service, so there is no app-store listing, review history or version number to check; the only evidence about it is on its own domain. Second, the last two rows are the two a buyer cares about most, and they are exactly the rows that third-party pages tend to fill in with numbers. If you see a monthly price or a count of free analyses quoted for PinkMirror anywhere, including on a competitor's page, check it against the pricing page itself before you rely on it. We did not, and could not, so we do not repeat one.

Most 'AI rating' tools are black boxes. PinkMirror is, to its credit, unusually open about the outline of its method, in a public post titled 'Face Beauty Research'. The method has four layers, and each has a known strength and a known weakness.

Layer 1: geometry from your photo

The post says the tool 'works by automatically extracting geometric measurements of your facial features from images'. That is the same first step every landmark-based analyser takes: locate points on the face, then compute distances, widths and angles between them. The strength is that this part is reproducible. The weakness is that a flat photograph is a two-dimensional projection of a three-dimensional face, so every distance depends on camera position, lens and head angle. The measurement is only as good as the photo that fed it.

Layer 2: three classical rules

The post names the rule of fifths (the face divides into five roughly eye-wide columns), the rule of thirds (hairline to brow, brow to nose base, nose base to chin, in equal parts) and a third it calls the 'Rule of Seventy Seven'. The first two are textbook proportion canons, and they are the oldest ideas in facial analysis. They are also much weaker predictors than their reputation: a critical analysis of one famous version, the 'phi mask', found that it describes fashion models and a mathematical ideal rather than what people actually rate as attractive (Holland 2008). The honest reading of these rules is as a description of balance, not a recipe for beauty.

Layer 3: feature tables fitted on a sample of faces

The most concrete part of the post is a set of tables. For each of roughly forty measurements, PinkMirror lists whether 'larger' or 'smaller' is better, an 'effectiveness' weight from 0 to 10 and an 'ideal' value, separately for female faces, male faces and all faces. The tables say they come from studying 'hundreds of faces'. Publishing the table is more transparent than most competitors manage, and it is the reason we can say anything specific here. It also has consequences you should understand before trusting a score built on it.

Read the first row slowly. The same measurement is better when larger in one table and better when smaller in another. The same row of the table also moves from a weight of 6.7 to 2.4 depending on which table you are in, and canthal tilt goes from 5.5 in the male table to 10 in the all-faces table. A score is therefore partly a function of which table the tool decides to use for your photo. That is not a flaw to hide; it is a property of any weighted-feature model. But it is why two tools, or the same tool on two photos, can disagree by a full point without either being broken.

Layer 4: a model that turns it into a number

The last step, in the post's words, is 'the machine's judgment, which was designed with the help of research and data of associated rating scores from different parts of the world'. This is the layer you cannot inspect: how the measurements are combined, what the training ratings were, who rated them. That is normal, and it is also the layer where the claim to be 'accurate' lives. Which brings us to the question people actually search.

Accuracy is three different questions that get used as one. Separating them is the single most useful thing you can do with any face-scoring tool, and it takes ten minutes.

Test 1: repeatability

Take five photos in one sitting: same light, same distance, same expression, head level. Score all five. A good tool returns nearly the same number five times. If it varies by more than half a point on a ten-point scale, the noise is larger than any change you could hope to see from a haircut, a new brow shape or a skincare routine, and re-testing after those changes tells you nothing. Anything the tool tells you is only meaningful if it is bigger than this noise.

Test 2: transparency

Can you see which measurements moved the score? PinkMirror's published tables answer part of this question, which is more than most tools do. The test to apply is stricter: after you get a number, can you point at the lines on your own photo and say which ratio is responsible? If not, you have a verdict without working. A number you cannot decompose cannot be acted on, because you will not know whether to change your hair, your photo or nothing at all.

Test 3: validity

Does the score predict what real people say about the face? This is the test almost nobody publishes, and it is the one that decides whether a score means anything. We ran it on our own tool and printed the result. On the 102 adults in the Face Research Lab London Set, the composite of symmetry and three ratios did not predict the average rating of 2,513 raters (ρ = −0.07). That is the honest state of the field: facial symmetry and averageness do influence attractiveness in the research literature (Grammer & Thornhill 1994; Langlois & Roggman 1990; Rhodes 2006), but the effects are modest, and a composite built from geometry alone does not recover human ratings. Ask any tool for its equivalent number. If there is none, the claim to accuracy has nothing under it.

PinkMirror's own accuracy line is more modest than the marketing you may have seen elsewhere: 'no analysis can claim absolute accuracy'. We agree with it, and we would go one step further. The question to ask is not whether the number is accurate, but what it is accurate about. A symmetry deviation or a facial-thirds ratio is accurate about your photo. A single beauty number is a convention.

The most common complaint about every face scorer is inconsistency, and the cause is almost always the photograph. The face did not change; the projection of it did. Five things move the numbers most.

1. 1Camera distance. A phone held at arm's length enlarges whatever is closest to the lens, so the nose looks wider and the jaw narrower than the same face shot from two metres with a zoom. Prop the phone, step back, zoom in.
2. 2Head angle. A chin lifted or dropped by a few degrees changes the thirds; a head turned a few degrees changes every left-right measurement, so a slight turn alone can produce a 'symmetry' result that is really a posture result.
3. 3Expression. A smile lifts the cheeks, shortens the eye opening and moves the mouth corners. Score a neutral face for structure, and treat a smiling photo as a different measurement.
4. 4Light. Side light puts a shadow across one cheek, and a landmark model that places a point on a shadow edge will read a narrower face on that side.
5. 5Mirror versus camera. You see yourself reversed in the mirror every day, and in one study of women having plastic surgery, 73% preferred a mirror-reversed photo of themselves to the standard one (de Runz 2016). A score read off an un-reversed photo can look 'wrong' to you before it is wrong.

Visible asymmetry is also less decisive than it feels. In a controlled study that altered one face in small steps, observers did not detect asymmetry until it reached about 3 mm (Chu 2011), which is why a tool that prints a symmetry percentage to one decimal place is reporting more precision than a viewer would ever perceive.

This is where we have to be plain about a gap. PinkMirror's menu lists a Pricing page. On 16 October 2026 it did not load for us, and nothing else on the site we could read states a price, a number of included analyses, or a renewal term. So this article contains no PinkMirror price. What follows is the checklist we would use before paying any beauty-scoring service, PinkMirror or otherwise.

For comparison, ours are public: paid features here run on credits, from a $4.99 pack with no subscription, or from $7.99 a month, and unused credits are refundable with no time limit under the refund policy. The measurements themselves are free and unlimited. That last point is the one that matters for re-testing, because the cost of a beauty score is almost entirely the cost of running it again.

It is a real, long-running site: its help page is dated 2016 and its research post 2020. Legitimate, though, is not the same as safe for a face photograph, and we cannot assess the second because we could not read its privacy policy. The question is the same for every analyser, and it has four parts.

• ✓How long is the photo kept? A number of days, or 'as long as necessary'? The second is not an answer.
• ✓Is it used to train models? Look for an explicit yes or no, and whether you can opt out.
• ✓How do you delete it? A button, an email address, a deadline in days. A face is biometric data in several jurisdictions.
• ✓Is there an age rule? Beauty-scoring tools are not made for children, and the research on looksmaxxing-style content already treats it as an emerging concern for teenagers (Mekhail 2026).

The strongest privacy position is also the simplest: a tool that measures in your browser never needs your photo at all. That is how the free measurements here work. Our AI features, including simulations and the written Face Report, do upload the photo you choose, only when you start one, and the privacy policy says which is which.

Most of what a score-based tool shows can be reproduced with free, single-purpose measurements, each of which tells you something you can act on. The table maps the pieces.

Run them on the same straight, level, evenly lit photo, then repeat after a change you care about. Because there is no per-analysis cost, the repeatability test from earlier is something you can do in five minutes rather than something you ration.

• ✓The measurement is free and unlimited. There is no count to ration and no account to open. Every re-test after a haircut, a brow change or a new routine costs nothing.
• ✓Your photo stays in your browser for the measurement. The landmark model runs on your device and the photo is not uploaded for it, so there is no retention window to read.
• ✓Every score has a reference group. Symmetry, fWHR, midface ratio and the composite are placed against 102 adults from the Face Research Lab London Set, measured with the same code. The distributions are published.
• ✓The formula is public, and so is its test. 478 landmarks, symmetry weighted 40% and three ratios 20% each, written out on the methodology page, with the validity result printed next to it, including the part that did not work (ρ = −0.07).
• ✓It names what a photo cannot measure. Each tool page says which traits need a profile view or depth, instead of guessing at them.
• ✓Pay per use, not per month. Credit packs need no subscription, and unused credits are refundable at the price paid.

A single photograph produces dozens of numbers, each with an 'ideal' beside it, and the human response to a list of ideals is to find a gap in every row. That is a design effect, not information about your face. Preoccupation with small or barely visible features is a recognised clinical warning sign: a systematic review and meta-analysis found body dysmorphic disorder markedly more common among people requesting cosmetic surgery (Salari 2022). If a score is making you anxious, stop scoring, keep the result private, and talk to a doctor or a therapist rather than to another tool.

Used well, measurements are useful in a narrow way: they tell you whether a change you made, a haircut, brows, posture, a photo set-up, moved something you can see. They are a baseline to improve from, not a rank. The sensible way to read any score, PinkMirror's or ours, is the way PinkMirror's own post ends: no analysis can claim absolute accuracy.