How AI Face Rating Works: 32 Vectors, PSL Calibration, and What We Actually Measure

Why AI for Face Rating?

Human raters are inconsistent. On a scale of 1–10, two experienced PSL raters typically agree to within ±1.2 points. Personal taste, halo effects (an attractive body inflating a face rating), photo quality biases, and rater fatigue all compound the variance.

AI doesn't have these problems. Once calibrated, a model applies the same criteria to every face at every time of day with zero fatigue, zero bias from secondary attributes, and consistent application of multi-dimensional metrics that humans can't compute reliably in their heads.

The 32-Vector Framework

Lookizm's model evaluates each face across four categories of measurement:

Category 1: Structural Vectors (10)

Bone-level proportions that define facial architecture:

• Bizygomatic width (cheekbone-to-cheekbone span)
• Bigonial width (jaw angle-to-jaw angle span)
• Bizygomatic-to-bigonial ratio (determines facial taper)
• Midface height-to-width ratio
• Mandibular projection (chin forward position)
• Gonial angle (sharpness of jaw corner)
• Anterior facial height (upper third, midface, lower third proportions)
• Facial thirds balance (forehead:midface:lower face)
• Facial fifths balance (eye width standardized across face width)
• Maxillary projection (midface forward depth)

Category 2: Harmonic Vectors (8)

Proportional and symmetry relationships:

• Facial symmetry index (horizontal and vertical axis)
• Inter-pupillary distance normalized to facial width
• Nose width-to-face width ratio
• Lip-to-nose vertical proportions
• Golden ratio alignment (phi ≈ 1.618 across key intervals)
• Eye-to-eyebrow distance ratio
• Nasal tip projection
• Lip vermilion-to-facial height ratio

Category 3: Dimorphism Vectors (8)

Sexual dimorphism markers — features that signal gender-typical hormone exposure and are closely tied to perceived dominance and attractiveness in male faces:

• Canthal tilt angle (medial-to-lateral canthus)
• Orbital depth (orbital rim to cornea projection)
• Supraorbital ridge prominence
• Brow position relative to orbital rim
• Lower lid exposure index
• Jawline angularity index
• Neck-to-jaw ratio
• Ramus height-to-body ratio

Category 4: Soft Tissue Vectors (6)

Surface-level quality markers that reflect genetic and lifestyle factors:

• Skin clarity and texture uniformity
• Subcutaneous fat distribution index
• Lip fullness and definition
• Periorbital area quality (under-eye, tear trough depth)
• Hairline position and density
• Overall skin tone evenness

PSL Calibration

Raw vector scores don't mean much in isolation. The model maps combinations of vector measurements to a PSL score by training on a dataset of over 9,000 faces that were rated by multiple expert human raters, with PSL scores established by consensus (3+ raters within 0.5 points of each other).

The resulting model predicts PSL with a median error of ±0.4 points against human expert consensus — tighter than two expert human raters typically agree with each other (±1.2 average disagreement).

Failos and Halos: Beyond the Number

The composite PSL score is useful but incomplete. What matters more for intervention planning is which specific vectors are driving the score up or down. This is what the failo/halo system captures.

Failos (architectural debits) are vectors where your measurement falls more than one standard deviation below the optimum for your tier. A recessed maxilla is a failo. A negative canthal tilt is a failo. These are the vectors where intervention has the highest expected PSL return.

Halos (architectural credits) are vectors significantly above average — these are what you build on and protect during looksmaxxing. A strong jawline is a halo. A low Norwood hairline is a halo. Don't accidentally harm your halos while fixing your failos.

What the Model Doesn't Measure

To be clear about limitations: the model doesn't factor in dynamic expression, voice, body language, height, body composition, or social behaviors. PSL rating is specifically a facial architecture metric. Off-face factors can add significant real-world social value beyond what a face rating captures — but they're outside the scope of a facial analysis tool.