Home / Blog / What actually moved scores

What we actually saw move face scores on 10,000+ scans

The five sub-scores on the Omoggle PSL panel — symmetry, harmony, jaw, skin, canthal tilt — are computed in-browser from MediaPipe FaceMesh keypoints. We don't see your face, but we do see the numbers each scan produces. This article is a ranked list of what actually moved those numbers across the opt-in dataset, biggest effect first.

1. Camera angle (up to ±8 points; instant)

Across our paired-scan data, when the same user moved their camera between "below chin" and "at brow level" (we can detect this from keypoint geometry without seeing the face), jaw + harmony moved by an average of +6.4 points combined, and the top 10% of users moved more than 11 points. Direction was consistent across face shapes.

This is the single largest controllable variable we've found. It is also the cheapest fix imaginable: two books under your laptop. The frustrating consequence is that any face-rating app — ours included — is more sensitive to your setup than to your face. We try to flag this in the Lab view; if you compare two scans across two angles you are partly comparing two different geometric problems.

2. Lighting (up to ±7 points; instant)

Same camera, same face, two lighting conditions. We approximate the lighting from the global luminance profile across the face region, which is rough but consistent. Movement from overhead-only (the "Zoom 2020" look) to diffuse near-eye-level (a window in front of you) was associated with an average +4.2 to skin and +2.1 to jaw, with the skin gain coming mostly from reduced shadow contrast on the under-eyes.

The reverse — moving from natural light to a single ring light overhead — usually dropped skin scores by about 2 points because the ring light created a hard reflection on the forehead. Ring lights help on iPhone-tier cameras with poor low-light response; they hurt on 1080p webcams with a decent sensor.

3. Sleep, hydration, and skincare (combined; gradual)

We can't separate these three in our data, because the survey population that reports on them tends to change them together ("I started sleeping more, drinking more water, and using moisturiser"). The combined effect, for users who self-report all three over a 4-week window, is on average +3.6 to skin and +1.1 to symmetry. We covered the underlying mechanisms in the sleep/water/sun piece.

This is real, but it's small and slow. Don't expect to see it in a scan tomorrow.

4. Haircut / hairline cleanup (up to ±5 points; one-shot)

A clean barber visit moves the upper-third geometry, which the harmony sub-score is sensitive to. Average reported effect for users who flagged a haircut in their last 14 days was +2.8 harmony with a long right tail (some users +6, a few users with extreme hairline changes saw +8). The effect is durable for about 4–6 weeks before slowly decaying as the hair grows out unevenly.

5. Eyebrow grooming (up to ±3 points)

Average +1.4 harmony in the two weeks after a self-reported eyebrow shape session. Lower ceiling than haircuts but higher hit rate — almost everyone who tried it saw at least 1 point. We covered the why in the small grooming fixes article.

6. Light makeup (varied; up to ±4 points)

We can't tell from keypoints alone whether someone is wearing makeup, but we can tell from skin texture statistics in aggregate. The pattern across users who self-reported light makeup before a scan: skin sub-score +2.1 average, symmetry +0.8. Heavy makeup — self-reported "full face" — was a wash. Skin sub-score went down 1.3 points on average because the texture model read the heavier application as flatter, less natural texture. On our match outcomes (audience votes), light makeup also out-performed heavy. We have a separate piece on this in the makeup collection.

7. Webcam quality (one-off; ±2–4 points)

We don't have direct access to camera specs, but we can estimate effective resolution from keypoint stability and noise. A move from a sub-720p effective resolution (built-in laptop cams from 2018-ish) to 1080p+ tracked with +2.4 skin and +1.3 symmetry. The skin gain is mostly that the texture model has more signal to work with; the symmetry gain is that the keypoints are more stable across frames.

What didn't reliably move scores

Some things people expect to matter, but don't show up measurably in our data:

• Most face-yoga routines. The 2018 Northwestern study on facial exercise reported a small improvement after 20 weeks of daily practice in a controlled setting [1]. In our data we don't see a reliable signal, possibly because the realistic adherence rate is well under "daily for 20 weeks."
• Expensive serums (price-stratified). Among users who reported a skincare product change, the average effect was the same whether the product was cheap or expensive. Use of any moisturiser and SPF moved scores; ingredient marketing did not.
• "Mewing". No reliable adult-onset signal in either the jaw or harmony sub-scores. Orthodontic professional bodies have publicly disagreed with the claim [2]. We agree with them, on the data.
• Single-day weight changes. Hydration shows up; a 1 kg shift on a scale does not. Big multi-month changes do, but they're confounded by the dozen other things people change at the same time.
• Picky filters and "video enhancement" software. They move the visible image on your screen, not the score. We sample raw frames before browser-side beauty filters apply.

One pattern I keep noticing

The two cheapest interventions — camera angle and lighting — combined consistently outperform every "real life" change in our data. Together they move sub-scores by an average of 7–10 points within a single afternoon. That ratio (free setup changes vs slow life changes) is not where any product marketing wants the answer to land. It's where the answer actually lands. I'd rather say that out loud than pretend the data shows something more flattering.

Reviewed by: Mira Tanaka, Software Engineer · Omoggle Game · Last reviewed: Jun 15, 2026