How AI Face Analysis Actually Works

Take a selfie, and within seconds an AI tells you which countries' people find your face most attractive? A few years ago this would have sounded like science fiction. Today it's a reality — but how exactly does it work? Most people have no idea what's happening under the hood when AI analyzes a face. In this post, we break down the technology in plain language, step by step.

Step 1: Mapping Your Face with 478 Points

The first thing AI face analysis does is locate key reference points on your face. Google's MediaPipe FaceLandmarker detects 478 landmark points from a single photo in real time. These points are distributed across every significant part of the face: the edges of the eyelids, the center of each iris, the arc of the eyebrows, the tip and nostrils of the nose, the outline of the lips, the contour of the jawline, and the position of the cheekbones.

By calculating the x, y, and z coordinates of all 478 points, the system builds a precise three-dimensional map of your face. This isn't just identifying positions in a flat 2D image — depth information is also estimated, so the system can tell whether you're looking straight ahead or slightly to the side. When GPU acceleration is available, this entire mapping process takes just milliseconds.

Step 2: Translating Your Face into 5 Key Metrics

Once the landmark coordinates are established, they're used to compute a set of core facial ratios. Here's what each one measures:

Face Ratio is the proportion of your face's width to its height. East Asian cultures tend to favor relatively narrow, elongated faces (a higher face ratio), while Mediterranean cultures often consider a wider, more balanced ratio ideal.

Jaw Width is the width of the jaw at its widest point divided by the overall face width. A lower number means a slim, tapered V-line jaw; a higher number indicates a broader, more angular jaw. Korean beauty ideals strongly favor low jaw width, while Western cultures tend to associate wider jaws with maturity and strength.

Eye Ratio measures the vertical height of the eye relative to the width of the face — essentially, how large your eyes appear in proportion to your face. Lip Ratio captures the combined thickness of the upper and lower lips, reflecting how full the lips are. Cheek Ratio measures the width at the cheekbone area, which has a major influence on overall face shape.

Step 3: Matching Your Metrics to Global Beauty Standards

With your five facial metrics calculated, the system compares them against beauty standard datasets for 139 countries. Each country has a defined range of ideal facial proportions that reflects its cultural aesthetic preferences. Korea, for example, favors a relatively high face ratio (elongated), low jaw width (slim), high eye ratio (large eyes), and moderately full lips. Each country's ideal profile is distinct.

Scores for each country are computed by measuring the distance between your actual metric values and that country's ideal range. The closer your measurements are to a country's ideal, the higher your score. Each metric is then weighted by its cultural significance — for instance, eyes carry a weight of 800 points, lips 450 points, and jawline 280 points — reflecting how prominently each feature factors into that culture's beauty judgments.

Step 4: Cultural Affinity Adjustment and Final Rankings

Raw metric comparisons alone can't fully capture cultural nuance. Countries within the same cultural sphere — say, Korea, Japan, China, and Taiwan — share overlapping beauty ideals. A face that scores highly in one of these countries will likely score well in the others too. To reflect this, the system applies a "cultural affinity" adjustment step.

Twelve cultural regions are defined: East Asia, Southeast Asia, South Asia, the Middle East, Western Europe, Eastern Europe, Southern Europe, Northern Europe, the Anglosphere, Latin America, Africa, and Central Asia. Within each region, scores are cross-referenced and calibrated to ensure that the final rankings meaningfully reflect genuine cultural preferences rather than statistical noise.

Finally, all country scores are normalized to a 0–100 scale and ranked. The top results are delivered to the user. Face detection and landmark extraction happen locally in your browser using MediaPipe, and only the numerical metric data is sent to the server for country matching. Your photo never leaves your device, keeping your privacy fully protected.

The Meaning — and the Limits — of AI Face Analysis

AI face analysis is a fascinating tool for exploring cultural diversity and what "attractive" means in different parts of the world. But it has important limitations worth understanding. First, photo quality matters: lighting, camera angle, and resolution all affect results. Second, the beauty standards baked into the system reflect a particular moment in time — beauty trends evolve, and the AI cannot always keep pace. Third, and perhaps most importantly, attractiveness is far more than facial geometry. Expression, energy, charisma, personality — none of these can be captured in a set of landmark coordinates.

That said, AI face analysis offers something genuinely new: a way to look at your own face through dozens of different cultural lenses at once, and to discover where on Earth your features are most celebrated. Somewhere out there, your face is exactly someone's ideal. Let's find out where.

Inside Hogamdo's Actual Face Analysis Pipeline

Beyond the general principles above, here is how Hogamdo actually implements the pipeline. Hogamdo classifies face embeddings from 139 countries into 13 cultural regions (EA, SEA, SA, ME, WE, EE, LAT, AFR, etc.) and computes a hogamdo (favorability) score from 11 quantitative facial metrics derived from the 478 MediaPipe landmarks.

The pipeline runs in five stages:

1. Landmark detection — MediaPipe FaceLandmarker extracts 478 3D points from the face.
2. Metric extraction — From those coordinates we compute 11 metrics: face ratio, jaw width, eye ratio, lip ratio, cheek width, and 6 supporting features.
3. Country matching — Cosine similarity is computed between the user's metrics and the 138-country embedding database.
4. Weight application — Each metric contributes differently (eye 800, lip 450, jaw 280, cheek 280, face ratio 30).
5. Two-tier normalization — Scores are mapped to 75–95 within the user's own cultural region and 65–100 across other regions.

Empirically, across our 13 regions, Latin America (LAT) accounts for ~16.4% of top-country results, the Middle East (ME) for 13.1%, Africa (AFR) for 11.5%, while East Asia (EA), Southeast Asia (SEA) and South Asia (SA) sit around 9%, and Eastern/Western Europe (EE/WE) around 7–8%. The top country in a result reflects the cultural region whose facial metric profile best matches the user's.

Limitations: Photo quality, angle, and lighting can shift results. These are statistical patterns, not absolute judgments of attractiveness. Use results for entertainment and cross-cultural curiosity only.