Digital Human

A digital human is a computer-generated person — a face, a body, a voice and increasingly a conversational mind — rendered in real time or offline so that it can speak, emote and respond in ways that read as human. The term covers a wide range: photoreal characters built to be indistinguishable from filmed actors, stylized avatars such as animated mascots and virtual singers, and AI-driven agents that greet customers, tutor students or play non-player characters in games. What unites them is a pipeline that must solve three separate problems at once: how the character looks, how it moves, and what it says and sounds like.

The visual problem is solved with scanned geometry, layered skin shading and a rig of controllable expressions, and the field's long-running benchmark is the uncanny valley. Masahiro Mori proposed the uncanny valley in 1970: as a figure approaches human likeness, affinity rises until a near-miss in skin, eyes or timing suddenly produces unease. Early landmarks such as the Digital Emily project (2008, USC Institute for Creative Technologies with Image Metrics) showed that a scanned actor could cross that valley in a controlled setting. Today Epic Games' MetaHuman framework for Unreal Engine packages that craft: characters with a full facial rig, hair strands and physically based skin are assembled from a library of scanned variation, then animated with the same real-time renderer used for games. Stylized digital humans — Hatsune Miku-style virtual singers, VTuber avatars, virtual influencers such as Lil Miquela, who appeared in 2016 — deliberately avoid photorealism, which sidesteps the valley and often reads as more trustworthy and expressive.

Voice and lip sync are the part where machine learning does the most visible work. Speech is converted into visemes, the mouth shapes that correspond to groups of sounds, and a model drives dozens of facial controls so that jaw, lips, tongue and brows move plausibly with the audio. Tools such as NVIDIA's Audio2Face generate facial animation directly from an audio track, while modern text-to-speech and voice-cloning models supply the voice itself. A conversational digital human chains these together: speech recognition turns the user's words into text, a large language model writes a reply, text-to-speech renders it, and the animation model lip-syncs the result, with each stage adding latency that must stay low enough to feel like conversation rather than a delayed broadcast.

Real-time rendering is the constraint that separates a digital human from a film visual effect. A cinematic character can spend minutes per frame on path-traced skin and hair; an interactive one has roughly sixteen milliseconds per frame at sixty frames per second, shared with the rest of the scene. That budget forces trade-offs — approximated subsurface scattering for skin, level-of-detail swapping for hair cards and strands, compressed animation, and sometimes streaming the rendered video from a server. The most expensive details are usually eyes, mouth interior and micro-expression timing, because human viewers are exquisitely tuned to those cues and notice a stilted blink or frozen gaze long before they notice a low-resolution pore.

The same technology that makes helpful digital humans also makes deceptive ones, which is why deepfakes sit at the center of the policy debate. The word "deepfake" emerged in 2017 from a community sharing face-swapped video made with neural networks, and the capabilities have since broadened to cloned voices and fully synthetic presenters. Documented harms include non-consensual intimate imagery, impersonation fraud — including reported cases of staff being tricked by synthetic video-call participants — and political disinformation. Responses are partial and still evolving: platform disclosure rules, content-provenance efforts such as the C2PA standard for signed metadata, detection research that is locked in an arms race with generators, and legislation such as the transparency obligations for synthetic media in the EU AI Act. Consent and likeness rights for actors, whose scans and voices can be reused, were also central issues in the 2023 Hollywood labor negotiations.

A frequent misconception is that a convincing face is the hard part; in practice, behavior is. A flawless render that gazes blankly, answers after a two-second pause or smiles at the wrong moment feels less human than a simple cartoon with well-timed reactions. Likewise a digital human is not an AI by itself — the face may be driven by a script, a human performer via motion capture, or a language model, and the model behind the voice may know nothing about what the character looks like. Evaluating a digital human therefore means asking separately about fidelity, responsiveness, truthfulness of the dialogue, and disclosure to the user that they are speaking with a synthetic character.

Read digital human as a stack rather than a face: scanned and rigged appearance, learned voice and animation, a language-driven mind and a real-time renderer, where believability depends on the weakest layer and trust depends on being honest that the person is synthetic.