In the late 1870s Francis Galton, a half-cousin of Charles Darwin with a lifelong compulsion to measure people, obtained a stack of prison photographs through Sir Edmund Du Cane, the man who ran the English convict system. He sorted out the men convicted of murder, manslaughter, or robbery with violence. Then he lined up their eyes and exposed their faces one after another, each one faintly, onto a single sensitized plate. He was hunting for the criminal type. He expected the face of crime itself to rise out of the noise, the way a signal rises when you average away the static.

What rose was a pleasant-looking man. "The features of the composites are much better looking than those of the components," Galton reported. "The special villainous irregularities in the latter have disappeared, and the common humanity that underlies them has prevailed." The plate had been asked for a murderer and had returned, in his own later phrasing, "not the criminal, but the man who is liable to fall into crime." Which is to say: almost anyone.

Galton's purposes were ugly. He coined the word eugenics, and he went on to composite tuberculosis patients and Jews in search of other types he considered inferior. But he recorded the accident honestly, and the accident outlived the project. In 1990 the psychologists Judith Langlois and Lori Roggman repeated it with computers, blending digitized faces in groups of two, four, eight, sixteen, and thirty-two. The more faces went into the blend, the more attractive people rated the result. They titled the paper "Attractive Faces Are Only Average."

The astronaut

This week a Guardian reporter sat in a nearly empty Beijing cinema and watched Sanxingdui: Future Past, the first fully AI-generated feature to open in Chinese theaters. The review noticed that the characters all ran with the same smooth, video-game gait. It also noticed that one of the leads, an astronaut called Wu Xingyan, "looks creepily like Tilly Norwood," the synthetic actress a London studio has cast in an upcoming feature called Misaligned.

Nobody has to have copied anybody for that to happen. Two faces built on opposite sides of the world, by different teams with different tools, arrived at the same address because both were standing near the middle of the same plate. A video model is Galton's apparatus at a scale he could not have pictured: millions of faces, each exposed faintly, onto one set of weights. Ask it for "a woman in her thirties" and it develops the plate. The villainous irregularities are gone. Common humanity prevails. The result looks lovely and could be cast as anyone, which in practice means she has already been cast as everyone.

The Korean director Yeon Sang-ho, who made a four-part spin-off of Train to Busan entirely with generative tools this autumn, described the problem with a bluntness Galton never managed. "AI is the sum of common patterns," he told the Korea Times, "so it cannot capture unique individuality." He added an observation about his own business. People spend fortunes to make films look real, he said, and then put famous actors everyone knows in front of the camera. Viewers do not want reality. They want a convincing illusion delivered by a face they can find.

The face you cannot find

The composite has a cost that does not show up on any single frame. In 1981 Leah Light and two colleagues published a short paper with an unkind title: "Why Attractive People Are Harder to Remember." Faces rated highly attractive were also rated more typical, and typical faces were recognized less reliably on a second viewing. The audience likes the average face and then loses it.

Six years later Gillian Rhodes, Susan Brennan, and Susan Carey ran the experiment in the other direction. They took line drawings of familiar faces and fed them through a program that measured each face's distance from an average face, then pushed every difference further out. People identified these caricatures faster than accurate drawings of the same faces. The anticaricatures, pulled toward the average, were slowest of all. Memory, it seems, files a face as a list of departures from the norm, and the more pronounced the departures, the faster the file opens.

Biology seems to have planned for this. In 2014 Michael Sheehan and Michael Nachman reported in Nature Communications that human facial traits are far more variable than other measurements of the body, and that the genetic regions behind them carry signs of selection for diversity. The likeliest reason is social. A species that lives by remembering who owes whom needs faces that can be told apart. Several hundred thousand years of evolution have been pushing human faces away from the mean. The composite pulls them back in a single pass.

Casting is choosing a departure

Lee Van Cleef's first film was High Noon, in which he played one of the gunmen waiting at the station and did not speak a line. He spent the next dozen years as a heavy, the man in the back of the saloon. Then Sergio Leone, a director who composed entire sequences around eyes and the creases beside them, built For a Few Dollars More around that narrow, hawkish face. Nothing about it was average. That was the whole qualification.

Casting directors have always understood that a face on screen has a job, and the job is to be found. The audience has to pick a character out of a wide shot across a cut or inside a crowd, in about the time it takes to blink. It does this by irregularity: the nose, the ears, the gap in the teeth, the brow that sits too low. A cast of composites is a cast of pleasant strangers who dissolve whenever the camera pulls back. When an AI short feels confusing, the blame usually lands on the script. Often the script is fine and the two leads simply share a face.

Kierkegaard once wrote that he wanted two words on his gravestone, "that individual," the category he spent his life defending against the crowd's average. Galton spent his life hunting the type and found, to his credit, that the type was nobody in particular. Every generation model runs on Galton's method. Every filmmaker who uses one has taken up Kierkegaard's job, whether or not anyone mentioned it in the interview.

Write the flaw

In Veo, Kling, Seedance, Runway, WAN, and Grok Imagine alike, words like "beautiful," "striking," "handsome," and "attractive" are instructions to add more faces to the plate. Neutral descriptions drift there too, just more slowly. "A man in his fifties, gray stubble, tired eyes" is a demographic. The model will fill in everything you left unsaid with the middle of the distribution, and the middle is where the face goes to disappear.

Describe departures instead. A nose broken once and set a few degrees to the left. One eyelid heavier than the other. A gap between the front teeth. A jaw slightly too long for the rest of the face. Ears that catch the backlight. Two or three of these, physical and asymmetric, will do more for a character than a paragraph of personality. Better still, build the character as a still first, choose the image for its flaw rather than its polish, and use it as the reference for everything that follows.

Then watch the flaw across the sequence. Reference and character-consistency features tend to hold the broad strokes of an identity while sanding off its small departures one generation at a time, so the scar that was obvious in shot two has faded by shot nine and the character has quietly regressed toward the plate. Check the irregularity the way a script supervisor checks a cigarette length. And before you lock a cut, shrink a few frames to the size of a postage stamp and put your leads side by side. If you cannot tell them apart at that size, neither can anyone watching the wide shot.

Galton wanted the face of a murderer and the plate gave him a neighbor. The lesson took a century to land, and the models have learned it perfectly. What survives on screen, what an audience finds in the dark and carries home, is the part the average removed.

Keep the villainous irregularities.


Bruce Belafonte is an AI filmmaker at Light Owl. He has asked for a nose broken in a bar fight eleven times and received eleven noses from a skincare commercial.