Let’s discuss some simple observations; if there are any errors, please feel free to correct me.
Human Factors
A friend mentioned above that human eyes are more sensitive to other humans. More specifically, our vision is particularly attuned to human posture and faces, a trait shared by most animals. For example, a group of chimpanzees can easily distinguish each other, while humans tend to see them as all looking the same. Chimpanzees are not actually identical—researchers who work with them for a long time can tell individuals apart—but the average person’s visual system pays insufficient attention to these differences and tends to “automatically ignore” them.
Inanimate objects and environments have similar reasons, for instance:
- Sensitivity to brightness: Humans are more sensitive to changes in luminance (brightness) than to changes in hue, so it is hard to discern details and quality in dark areas of a scene.
- Visual field limitation: Although binocular vision can cover over a hundred degrees, the truly focused field is very small (probably less than ten degrees). Attention and resolution are strongest at the center and weaker toward the edges. Therefore, areas that are not being directly looked at are unlikely to reveal poor quality, even if it is present.
Technical Factors
This topic is a natural extension of the characteristics of human vision—since our eyes notice many facial details, rendering a face often requires accounting for numerous complex biological features and optical effects. Below are a few examples to illustrate the point.
Skin Surface Details
Humans are not marble statues; no one’s skin is flawless. Scars, pores, fine hairs, and countless other details vary greatly, and modeling them is extremely difficult. The usual PBR approach that describes roughness and metallicity is insufficient for skin, so a large number of special maps are needed.
From a geometric scale perspective, these features are tiny and hard to approximate. Real‑time rendering pipelines involve many implicit sampling steps that are unfriendly to high‑frequency signals, making the rendering cost high. Visually, the presence or absence of these features can be decisive in whether a face falls into the uncanny valley.
Subsurface Scattering
This is the famous optical effect. If you place your finger over a phone’s flashlight, you’ll see a semi‑transparent glow around the edges. This translucency differs from that of a typical transparent object; it is modeled as a thin layer of semi‑transparent scattering medium on an otherwise opaque surface, requiring specialized rendering techniques.
Snow and similar materials also exhibit subsurface scattering, but snow has uniform color and few details. Skin, on the other hand, is completely different, making the combination of subsurface scattering and fine skin detail especially challenging.
Small‑Scale Asymmetry and Detail Motion
No human body is perfectly symmetrical, especially the face. This asymmetry is the kind that “you don’t notice it when it’s there, but it feels odd when it’s missing,” making it very hard for artists to model and for real‑time game rendering to allocate enough geometry for such subtle variations.
Moreover, facial muscles are extremely complex—any animal with facial musculature tends to be intricate, and we are especially sensitive to human faces. Current mesh deformation systems are not driven by muscle dynamics; they rely on large motion‑capture datasets supplemented with manual modeling. Motion capture can be very precise, but the number of usable vertices on a face is limited, making it difficult to reproduce subtle micro‑expressions. This is why game characters sometimes look unreal, yet it’s hard to pinpoint exactly what feels off.
In the professional digital‑human field, these issues can be partially solved, but the cost may be a single RTX 4090 capable of rendering only one face, which is clearly impractical for games today.
My understanding of rendering is rather shallow; the points above are well‑trodden topics that have been discussed for over a decade. Yet even now, many studies on digital humans still grapple with the challenge of achieving realistic human rendering. When all these challenges are placed into real‑time rendering, the difficulty jumps another level; put them into games, and the difficulty rises yet again. I wonder, ten years from now, when we look at games, will we see the level of digital humans that Jensen Huang talked about a couple of years ago?