top of page

The Uncanny Valley: Why Digital Humans Feel Eerie and How Studios Cross It

Mimic Productions
4 days ago
7 min read
Photoreal digital human face on a studio monitor, half finished skin and half wireframe, watched by an artist. Text: The Uncanny Valley

The uncanny valley is the drop in comfort people feel when something looks almost, but not quite, human.


A cartoon character feels friendly. A real person feels normal. Somewhere between the two sits a zone where a face is realistic enough to be judged as human, yet slightly wrong in a way viewers struggle to name. The eyes look glassy, the skin looks waxy, or the smile arrives a fraction too late. The result is unease rather than connection.


For anyone building a digital human, an AI avatar or a digital double, the uncanny valley is not a curiosity. It decides whether an audience trusts the character, listens to it and remembers the brand behind it.


This guide explains where the idea came from, what research says about why it happens, which technical details trigger it and how production studios design characters that cross it.



Table of Contents




What Is the Uncanny Valley?


Grayscale illustration of a curve rising from a simple robot to a stylised character, dipping sharply at a near-human mannequin face, then rising again to a real human portrait.

The uncanny valley describes a relationship between how human something looks and how comfortable people feel around it.


As a figure becomes more humanlike, our affinity for it usually rises. A simple robot is neutral, a stylised animated hero is likeable. But just before full realism, the curve falls into a dip. A face that is 95 percent convincing can feel colder than one that is 60 percent convincing, because our attention shifts from what is right to what is wrong.


Climbing out of the valley means reaching a level of realism where nothing contradicts the impression of a living person. That is the target for photoreal digital humans, and it is why small details carry so much weight.



Where the Idea Came From


Grayscale illustration of a 1970s Japanese engineer sketching a curve on paper, with a robot arm, a prosthetic hand and a traditional puppet on the desk.

The concept comes from Japanese roboticist Masahiro Mori, who published a short essay called "Bukimi no Tani" in the journal Energy in 1970. The title is often translated as "valley of eeriness."


Mori plotted human likeness against a feeling he called shinwakan, usually translated as affinity. His examples included industrial robots, toy robots, Bunraku puppets and realistic prosthetic hands, which can feel unsettling once we notice they are not living tissue.


Two points from the essay still matter for production today:


  • Movement deepens the valley. Mori argued that motion amplifies both the peaks and the dip. A slightly odd face is worse when it moves.

  • Aim deliberately. He advised designers to target the first peak, a clearly non-human but likeable design, rather than risk falling into the dip.


The essay received little attention for decades. An authorised English translation by Karl MacDorman and Norri Kageki appeared in IEEE Robotics and Automation Magazine in 2012, and Mori himself described the curve as guidance for designers rather than a proven scientific law.



Why the Uncanny Valley Happens


Grayscale illustration of a human brain beside a face where one eye is slightly larger and the skin is unnaturally smooth, with thin dotted lines linking it to the mismatched features.

Researchers have tested several explanations. The two most discussed are:


  • Categorisation difficulty: the brain struggles to decide whether the figure is human or not, and that ambiguity feels uncomfortable

  • Perceptual mismatch: some features look fully human while others do not, such as realistic skin paired with artificially large eyes, and the inconsistency triggers unease


A 2015 review by Kätsyri and colleagues in Frontiers in Psychology found good support for perceptual mismatch, too little data to settle the categorisation idea and no consistent support for the simple curve as Mori drew it. Later reviews reached similar conclusions.


For studios, this is the most useful finding. The valley is not triggered by realism itself. It is triggered by uneven realism, where one element of a character is much more convincing than another. That is a problem a production pipeline can control. Our article on the psychology of digital humans looks at how audiences read faces in more depth.



What Triggers It in Digital Humans


Grayscale close-up of a digital face with highlighted problem zones: glassy eyes, waxy cheek, stiff mouth corner and hairline.

In production, the same handful of details cause most uncanny reactions:


  • Eyes: missing moisture, flat reflections, no small eye movements and a lack of redness or veins at the edges

  • Skin: surfaces that look like plastic or wax because light does not scatter beneath them, a problem solved with subsurface scattering

  • Facial movement: expressions that switch on and off instead of building, or wrinkles that do not appear when muscles contract

  • Lip sync: mouth shapes that lag or lead the voice by even a few frames

  • Timing: blinks, breathing and micro-movements that are too regular or missing entirely

  • Hair and edges: hairlines, eyelashes and lip edges that look cut out rather than grown


None of these is dramatic on its own. Together they tell the viewer that something behind the face is not alive.



Famous Uncanny Valley Examples


Grayscale illustration of a cinema screen showing a row of computer-generated faces from different eras, from early stiff faces to a near-perfect modern face.

Film history offers clear lessons:


  • Final Fantasy: The Spirits Within (2001): one of the first fully photoreal CG features. Critics praised the ambition but noted cold eyes and mechanical movement.

  • The Polar Express (2004): built with performance capture, it is still the textbook example. Many viewers found the characters' eyes lifeless.

  • The Curious Case of Benjamin Button (2008): widely seen as a breakthrough. The aged digital head driven by Brad Pitt's performance won the Academy Award for visual effects.

  • Rogue One (2016): the digital recreation of Grand Moff Tarkin divided audiences, especially in scenes next to live actors where small differences were easy to spot.


The pattern is consistent. Projects that captured real performance detail and matched every layer of realism fared better than those where one layer lagged behind. Our guide to de-aging in film and television covers how this has improved since.



How Studios Cross the Uncanny Valley


Grayscale illustration of a pipeline from a face scanning rig, to a detailed facial rig with control points, to an actor in a head-mounted camera, to a finished realistic face.

Crossing the valley is less about one technology and more about keeping every layer of the character at the same level of truth.


A professional pipeline usually includes:


  • Scanning the real person: photogrammetry and expression scans capture true proportions, pores and wrinkles

  • Skin and eye shading: layered skin shaders with subsurface scattering, plus separate eye models for cornea, iris and moisture

  • A detailed facial rig: facial rigging built on FACS expressions and corrective blend shapes, so wrinkles and skin folds follow the muscles

  • Real performance data: performance capture or 4D facial capture preserves the timing of real expressions

  • Hair, lashes and brows: groomed strands rather than cards wherever the camera is close

  • Animation polish: artists refine eyes, blinks and lip contacts by hand, because these are where audiences look first


The final step is review in context. A face can look perfect in a turntable and still fail in a scene with real lighting, live actors or a moving camera.



Real-Time Avatars and the Moving Valley


Grayscale illustration of a person talking to a realistic digital human on a large screen in a lobby, with a small sound wave between them.

Real-time digital humans and conversational AI avatars face a harder version of the problem. They must look right, and also respond right, live.


Extra risks include:


  • Latency: a pause before the avatar reacts reads as unnatural, even if the face is perfect

  • Generated lip sync: speech-driven animation can drift when the voice changes pace

  • Idle behaviour: an avatar that freezes between answers breaks the illusion faster than any texture issue

  • Engine budgets: real-time skin, hair and eyes must be optimised without losing the details that keep them believable


There is also a design choice. Not every product needs a photoreal human. For some brands a stylised character sitting on Mori's first peak is safer and just as engaging. Our comparison of AI avatars and digital humans explains how to choose, and real-time facial motion capture shows how live performance can drive a character.



Comparison Table


Character style

Uncanny valley risk

Main reason

Best suited to

Stylised cartoon

Low

Clearly non-human, no false promise of realism

Games, children's content, mascots

Semi-realistic

Medium

Realistic skin with exaggerated features can mismatch

Games, apps, virtual influencers

Photoreal, uneven quality

High

One layer, often eyes or motion, lags behind the rest

Avoid

Photoreal digital double

Low when complete

Scan, rig and performance all match a real person

Film, advertising, digital humans

Real-time conversational avatar

Medium to high

Latency, generated lip sync and idle behaviour

Customer experience, events, AI products



How Mimic Productions Approaches Realism


Photoreal digital human Leya created by Mimic Productions, a woman with freckles, green eyes and tied-back hair in a burgundy top against a grey studio background.

At Mimic Productions in Berlin, avoiding the uncanny valley starts with the brief. The first question is not how realistic a character can be, but how realistic it needs to be for its audience, platform and camera distance.


When photorealism is the goal, the team keeps every layer at the same standard: scanning, skin and eye shading, groom, facial rigging and performance capture are planned together rather than added one by one. The Leya avatar shown above is one example of this approach, built from 3D scanning through to a finished digital character.


The same principles apply to photoreal 3D character models for film and advertising and to real-time avatars for interactive products, where performance and responsiveness matter as much as looks.



Frequently Asked Questions


It is the uneasy feeling people get when a robot or digital character looks almost human but not quite. Small flaws become noticeable and make the figure feel eerie rather than friendly.

Japanese roboticist Masahiro Mori described it in a 1970 essay in the journal Energy. An authorised English translation was published in IEEE Robotics and Automation Magazine in 2012.

Research supports parts of it. Studies find that mismatched realism, such as lifelike skin with unnatural eyes, reliably causes discomfort, but the exact curve Mori drew is not consistently reproduced.

Usually because one element lags behind the others. Common causes are lifeless eyes, waxy skin, delayed lip sync, stiff expressions and missing micro-movements such as blinks and breathing.

The Polar Express and Final Fantasy: The Spirits Within are classic examples. The digital Grand Moff Tarkin in Rogue One divided audiences, while The Curious Case of Benjamin Button is often cited as a success.

They scan real people, use physically based skin and eye shading, build detailed facial rigs, drive faces with real performance capture and polish eyes and lip sync by hand.

No. If a project cannot support full realism on its platform or budget, a stylised design is often safer and more engaging than a near-real character that falls into the valley.

Yes. Real-time avatars add risks such as response latency, generated lip sync and frozen idle states, so behaviour matters as much as visual quality.



Conclusion


The uncanny valley is not a wall that blocks realistic digital humans. It is a warning about inconsistency.


Mori's curve and the research that followed point to the same lesson: audiences forgive a stylised face, and they accept a truly convincing one, but they reject a face where the eyes, skin, motion and timing disagree. Crossing the valley means planning every layer to the same standard and testing the character in the context where people will actually see it.


To discuss a realistic digital human, digital double or AI avatar, explore our photoreal 3D character models or contact Mimic Productions.



For inquiries, please contact: Press Department, Mimic Productions info@mimicproductions.com

Comments


bottom of page