Two tools can analyze the same face and display different numbers without either having made an arithmetic mistake. They may have located different points, measured different properties, or attached different meanings to the same scale.
The interface often hides those choices behind a single score. To understand a disagreement, follow the information from the photograph to the final display. That is more useful than assuming that the higher number is flattering or the lower one is honest.
This article was prepared for Looksmax Scale, whose tool is discussed below. Its implementation provides a concrete example of the choices involved. No comparison testing of competing products is presented here; the aim is to explain what a meaningful comparison would need to establish.
1. Detection and measurement are different stages
A system first needs to locate a face and estimate the features it will analyze. Those estimates can depend on visibility, pose, image quality, and the model used. The next stage turns the estimated positions into distances, angles, or ratios.
Google's Face Landmarker documentation describes a model that estimates 478 three-dimensional landmarks. Landmark prediction supplies facial geometry information; it does not itself establish an attractiveness rating.
That distinction matters because a product can use a learned model for one stage and fixed arithmetic for another. “AI-powered” does not tell you whether the final score comes from a trained rating predictor, a written formula, or a language model asked to evaluate an image.
Looksmax Scale's photo analysis uses estimated landmarks and calculates ratios from their projected horizontal and vertical positions after correcting for image aspect ratio. The resulting measurements describe the image representation. They are not equivalent to a calibrated three-dimensional scan, even though the underlying landmark model has depth-related outputs.
Imagine a mouth corner estimated a little higher in one analysis. A deterministic formula can faithfully turn that estimate into a changed result. The arithmetic can be stable while the input measurement remains imperfect. Debugging only the final number would miss the earlier source of variation.
Estimated landmarks provide inputs for later measurements; they are not attractiveness labels. Selected schematic points, not the complete 478-point model.
2. Scoring rules decide what counts
Once features have been measured, a product still needs a rule connecting measurements to a rating. It may choose particular target ratios, combine multiple deviations, and adjust the output to fit a display scale. Those are substantive design decisions.
For example, the Looksmax Scale attractiveness test combines four dimensions: harmony at 32%, symmetry at 26%, eye area at 21%, and jawline at 21%. Later transformations also affect the displayed score. The weights describe the combination of dimensions, but they do not fully specify the displayed result.
Another implementation could emphasize a different feature, normalize measurements differently, or use a different mapping from raw results to numbers. You should expect these choices to matter. Their existence does not establish which design better matches human judgments.
Here is a worked example using invented values and deliberately simple formulas. These are not Looksmax Scale's formulas or measurements of competing tools. Suppose both systems locate the same feature pair and find a gap of 6 pixels. They also agree that cheek width is 200 pixels and eye-center distance is 80 pixels.
| Step | Hypothetical system A | Hypothetical system B |
|---|---|---|
| Reference distance | Cheek width: 200 pixels | Eye-center distance: 80 pixels |
| Relative gap | 6 / 200 = 0.03 | 6 / 80 = 0.075 |
| Invented display rule | 10 × (1 − relative gap) | 10 × (1 − relative gap) |
| Displayed result | 9.7 | 9.25 |
Both systems used identical point locations and correct arithmetic. The disagreement entered when they chose different reference distances. The larger number therefore tells you which formula produced the smaller normalized gap, not which tool measured the face more accurately.
Documentation should therefore answer two questions: what was measured, and how did that become the number on screen? A list of facial landmarks answers only the first part of the first question.
Several separate design choices sit between the image and the displayed score.
3. A shared scale does not create a shared benchmark
A score of eight has meaning only inside a defined system. It might be a transformed geometric index, an estimate of average ratings from a labeled dataset, or a generated assessment without a documented numerical calibration.
You cannot convert these into equivalent measurements simply by matching their minimum and maximum. A zero-to-eight display and a zero-to-ten display may use different curves, categories, or reference assumptions. Multiplying one result by a constant does not solve that problem.
Continue the invented example with the normalized gap fixed at 0.03. The linear rule above produces 9.7. A different hypothetical rule, 10 × (1 − relative gap)², produces 9.409, or 9.41 rounded. Nothing about the photo or point locations changed. Only the display mapping changed. These toy rules illustrate a dependency; neither has been validated as an attractiveness measure.
Normalization and display mapping are useful places to look because an interface may conceal both. A product could document the landmarks while leaving its reference distance, feature weights, and final curve unexplained. Identical labels on the results would then offer little basis for a direct comparison.
Percentiles require particular care. A position within a specified reference sample is a claim about that sample and the measured variable. It does not automatically describe a person's position among all people in attractiveness. Ask who supplied the reference data and whether the output was evaluated against the claim being made.
The same caution applies to updates. If a product changes its score mapping, old and new numbers may not form a comparable sequence. Record the date and version when available. Otherwise, a software revision can look like a change in the person being photographed.
4. Repeatability depends on what you hold constant
There are several different questions hidden in “Does it give consistent results?” Does an unchanged file get the same score? Do different photos of one person get similar scores? Does the same photo produce similar results across devices? Each requires its own test.
Fixed arithmetic produces the same calculation from the same landmarks under the same code. Looksmax Scale is designed to return the same score for an unchanged image under the same implementation. That statement is narrower than a guarantee across browsers, model versions, or future updates. Cross-environment repeatability was not newly measured for this article.
Different photographs introduce additional variation. Moving the camera changes perspective; turning the head changes the visible projection; expression changes landmark positions. A result that moves between such pictures does not, by itself, reveal whether the formula is unstable.
A system using generated evaluations may introduce other dependencies, such as model version, instructions, and generation settings. The relevant question is how that particular product is configured and tested. The label “generative” alone does not provide a measured estimate of its variability, just as “fixed formula” does not provide evidence of accuracy.
Repeatability tests need to state which inputs and conditions stayed fixed. A test plan, not measured results.
5. Use a diagnostic checklist before choosing a winner
Start with the original file. Confirm that both systems received the same image rather than a screenshot, crop, or compressed copy. Record any automatic cropping or preprocessing you can observe.
Next, identify the output. Is it a symmetry index, a face-shape category, or a broader attractiveness rating? Check the scale and look for a definition. If those details are absent, keep the result labeled as poorly documented rather than filling in the missing explanation yourself.
Then separate a repeatability check from a photo-condition check. Reanalyzing one file addresses the first question. Comparing different poses addresses another. Keep a simple table of file, tool version, output type, and result so those questions do not merge.
Add a column for the first stage you can verify. If the tools expose point overlays, compare those before comparing their scores. If they expose raw ratios, compare the definitions and denominators. If they expose only final numbers, record that the source of disagreement remains unresolved. This prevents an undocumented calculation from quietly becoming your preferred explanation.
Finally, ask what independent validation exists for the intended use. Predicting a dataset's ratings and measuring a geometric ratio are different objectives. Neither alone proves that a tool can predict dating outcomes or summarize a person's appeal in daily life.
When two numbers disagree, the useful outcome is an explanation of their assumptions. If you can identify the measurement, the transformation, and the tested limits, you have learned something about the tools. A larger number on its own tells you much less.
Top comments (0)