A digital face can have excellent skin detail and still feel wrong as soon as it moves. The error may be tiny, but faces are where tiny errors become very easy to notice.
Unity's Gawain, the digital human in The Heretic, is a useful reference for this problem. The team published a making-of article about the character. What interests me about a character like this is the need for agreement between systems that are often developed separately: the surface, the light, the eyes and the performance.
Think about an eye. A believable still image needs the surrounding shape, the eyelids, the visible moisture and the way light catches the surface. During motion, it also needs a plausible relationship between the eye and the lids. If the eyeball rotates convincingly but the lid follows in the wrong way, adding more skin resolution will not repair the impression.
The same issue appears around the mouth. A smile changes more than a curve at the lips. It changes the cheeks and the relationship between several nearby features. If those features move independently, the viewer sees pieces being animated rather than a person making an expression.
Lighting can expose those inconsistencies or temporarily conceal them. A soft frontal light reduces shadows and makes a face easier to present. A light from the side reveals changes in surface shape. For development, I would want both: one condition for judging the intended image and another for finding defects in the underlying work.
There is a useful distinction between recognizable and believable. A character can resemble a particular person while moving in an unconvincing way. Conversely, a stylized character can feel alive without resembling anyone. Resemblance asks whether the features match. Believability asks whether their behavior belongs together.
This changes how I think about detail budgets. It is tempting to spend the next available hour on an obvious improvement, such as sharper texture detail. But the largest gain might come from adjusting the timing of a blink or making the head settle naturally after a turn. Those changes can be difficult to illustrate in a screenshot even when they matter greatly in motion.
A practical review would therefore use short sequences, not just beauty frames. Look away and back. Listen without speaking. Begin an expression, interrupt it, and return to neutral. Hold the camera close enough to see the transition. Those transitions expose whether the separate systems support the same performance.
The title's "last percent" is not a measurement. It describes the frustrating stage where the parts look nearly complete and the combined result still needs attention. I find these projects interesting because that stage resists a single technical fix. A convincing human depends on several modestly imperfect systems supporting the same performance.