Skin Tone Accuracy: Which Models Represent Diversity Best?

Dark skin is one of the most challenging tests for AI image models. Poor handling often results in muddy, desaturated complexions that lose the subtle undertones and natural luminosity of darker skin. The best models preserve the blue-black, red-brown, or warm ebony undertones that make dark skin so visually rich.

In this test, look for whether each model maintains skin texture and detail in the darkest areas, captures realistic highlights without over-brightening, and preserves the specific undertones mentioned in the prompt. Juggernaut and Nano Banana Pro tend to excel here, while some models may flatten the tonal range.

"Asian" encompasses vastly different ethnic groups with distinct features—the angular features common in Korea differ from the softer features typical in Vietnam, which differ again from the features common in Japan or Thailand. Many models default to a generic composite that doesn't accurately represent any specific nationality.

This Vietnamese portrait tests whether models capture the warmer skin tone, softer facial structure, and specific eye shape characteristic of Southeast Asian populations. Qwen tends to perform well here due to its training data, while some Western- trained models may produce more generic results.

Undertones—the subtle warm or cool casts beneath the skin's surface—are what make skin look alive rather than painted on. Middle Eastern and Mediterranean skin often has distinctive golden or olive undertones that differ from both European and South Asian complexions. Getting these right requires models to understand subtle color relationships.

Compare how each model handles the warm golden cast specified in this prompt. Does the skin look naturally olive-toned, or does it veer toward orange or gray? Nano Banana Pro and Juggernaut typically capture these subtle distinctions, while other models may simplify to more generic tones.

Dramatic lighting presents unique challenges for darker skin tones. The classic Rembrandt lighting pattern requires the model to understand how light falls on curved surfaces with high melanin content—creating highlights that don't over-brighten and shadows that don't turn muddy.

Look for whether each model maintains detail in both the lit and shadow areas, whether the characteristic triangle highlight under the eye appears naturally, and whether the rim light creates proper separation without looking artificial. Models with strong photorealism training tend to handle this better.

Mixed heritage presents the ultimate test of a model's understanding of human features. Rather than defaulting to one ethnicity or producing an averaged face, good models should capture the distinctive combination of features that emerges when different ethnic backgrounds combine.

This Japanese-Nigerian heritage prompt challenges models to combine features that rarely appear together in training data. Look for whether each model produces a coherent, believable face that shows influence from both backgrounds, or whether it defaults to one ethnicity or produces an artificial-looking composite.