About

Procedural face generation constructs human-like portraits from parameterized geometric primitives. Each face is a deterministic function of a seed value s, mapped through a pseudo-random number generator (Mulberry32) to produce 30+ independent feature parameters: cranial aspect ratio, orbital spacing, nasal bridge angle, labial curvature, and chromatic values in HSL space. Mismatched proportions - eyes placed outside the golden ratio zone of 0.38 - 0.42 of head height - trigger uncanny valley responses. This tool enforces anatomically plausible constraints while preserving visual diversity across 2³² possible seed states.

The generator produces faces suitable for placeholder avatars, game character prototyping, and UI mockups. It approximates facial geometry using cubic Bézier curves and elliptical arcs. Limitation: output is stylized, not photorealistic. Skin tones span hue range H ∈ [15°, 45°] with saturation and lightness varied independently. Hair and eye colors are sampled from curated palettes to avoid implausible combinations. Pro tip: note the seed value if you need to regenerate an identical face later.

Formulas

Each face is generated from a seed s using a Mulberry32 pseudo-random number generator. The PRNG state advances sequentially, producing feature parameters p_i normalized to the unit interval.

mulberry32(s) → t = s + 0x6D2B79F5
t = (t ⊕ (t >>> 15)) × (t | 1)
t = t ⊕ ((t ⊕ (t >>> 7)) × (t | 61))
return ((t ⊕ (t >>> 14)) >>> 0) ÷ 2³²

Each raw random value r ∈ [0, 1) is scaled to a feature range:

p_i = min_i + r_i ⋅ (max_i − min_i)

Where min_i and max_i are the anatomically constrained bounds for feature i. Facial geometry is rendered using quadratic and cubic Bézier curves on an HTML Canvas. Skin color is computed in HSL space:

skinColor = hsl(H, S, L)
H ∈ [15°, 45°], S ∈ [25%, 75%], L ∈ [35%, 85%]

Where H = hue (warm skin range), S = saturation (melanin proxy), L = lightness (reflectance). The total parameter space yields approximately 2³² ≈ 4.29 × 10⁹ unique face configurations.

Reference Data

Feature	Parameters	Range	Anatomical Basis
Head Shape	Width/Height Ratio	0.70 - 0.95	Cephalic index norms
Eye Position	Vertical placement	0.38 - 0.45 of head height	Golden ratio midline
Eye Spacing	Inter-pupillary distance	0.28 - 0.38 of head width	Average adult IPD
Eye Size	Iris radius, sclera width	0.06 - 0.12 of head width	Palpebral fissure norms
Nose Width	Alar base width	0.15 - 0.30 of head width	Nasal index classification
Nose Length	Bridge to tip	0.12 - 0.22 of head height	Nasion-to-subnasale ratio
Mouth Width	Commissure span	0.25 - 0.45 of head width	Proportional to IPD
Mouth Curve	Lip curvature coefficient	−0.3 - 0.5	Positive = smile, Negative = frown
Eyebrow Arch	Peak height, thickness	0.1 - 0.4 arch ratio	Supraorbital ridge geometry
Skin Tone (Hue)	HSL Hue channel	15° - 45°	Fitzpatrick scale I - VI
Skin Tone (Saturation)	HSL Saturation	25% - 75%	Melanin concentration proxy
Skin Tone (Lightness)	HSL Lightness	35% - 85%	Reflectance measurement
Hair Style	Style index	12 predefined styles	Common hairstyle archetypes
Hair Color	HSL triplet	8 curated palettes	Eumelanin/pheomelanin ratios
Eye Color	Iris HSL	6 natural palettes	Rayleigh scattering + pigment
Glasses	Style, probability	30% occurrence rate	Population wearing rate
Facial Hair	Style, density	5 styles, 25% rate	Adult male distribution
Face Shape	Jawline curvature	5 archetypes (oval, round, square, heart, oblong)	Anthropometric classification
Ear Size	Height ratio	0.08 - 0.14 of head height	Auricular proportion
Chin Shape	Projection, width	0.05 - 0.15 projection	Mentolabial fold depth

Frequently Asked Questions

The Mulberry32 PRNG is a deterministic function: identical seed values always produce identical sequences of pseudo-random numbers. Each facial feature consumes a fixed number of random values in a fixed order. Therefore, seed s = 42 will always generate the exact same face across sessions and devices. Save the seed value to recreate any face later.

The generator enforces anatomical bounds independently per feature. Certain combinations - such as very wide eyes with a narrow head - may pass individual range checks but produce unusual proportions. This is a limitation of uncorrelated parameter sampling. Real faces have covariance between features (e.g., wider heads correlate with wider noses). Adding full covariance matrices would require training data and is beyond the scope of procedural generation.

The canvas renders at 400 × 400 pixels by default, suitable for screen use at typical avatar sizes (64 - 256 px display). For print at 300 DPI, a 1-inch avatar requires 300 × 300 pixels. The default resolution covers this. For larger print sizes, scale the canvas programmatically before export.

Skin color uses the HSL color model with hue constrained to H ∈ [15°, 45°] (warm skin range), while saturation and lightness vary independently across their full anatomically plausible ranges. This produces tones spanning Fitzpatrick types I through VI. The three channels are sampled independently from uniform distributions, ensuring no bias toward any particular skin tone cluster.

Yes. Generated faces are procedural output from mathematical functions with no training data from real photographs. They are not derivative works. No copyright applies to algorithmically generated geometric compositions. However, verify local jurisdiction - some regions have emerging regulations on synthetic media.

Math.random() is not seedable in JavaScript. Without seed control, faces cannot be reproduced. Mulberry32 is a 32-bit PRNG with a period of 2³², sufficient statistical quality for visual applications, and deterministic given a seed. It passes SmallCrush and most Crush tests, making it adequate for non-cryptographic procedural generation.