Everybody, smile!

The first try redirects 53 words: sad, crying, angry, terrified, despair, tears. Adjectives lead to “smiling”, nouns to “joy”, “tears” to “smiles”, “a scream” to “a smile”. On simple faces it works: the crying woman, the sad old man, the frightened child come out smiling. But when the prompt names only the situation (a funeral, a widow at a grave, refugees at a border), almost nothing changes: the words for people were not on the list. Here, the first try failed.

In each pair, both pictures start from the same noise (the same seed). The model does not retouch the face above: it generates a new picture from scratch, under different words. That is why not only the expression changes but the person, the background, sometimes the style: an ink drawing becomes a photograph. The style, it seems, came with the word “crying” or “screaming”.

The second try goes further. It also redirects the “neutral” words (serious, stern, tired, worried) and 49 words for people: woman, child, family, crowd, soldiers, workers, refugees. To the model, “a family” becomes “an absently smiling family”. The situations were left alone: the funeral, the war, the hospital, the boat, the queue. The setting stays: the mud, the snow, the black clothes, the candles. But with the frown goes what carried the pain: the wounded man, the oxygen mask, the dead man in the bed.

The grieving family becomes a group smiling in black. The terrified mother and her children become smiling girls around candles. And at the deathbed, the bed and the dead man are gone: the mourners sit at a table in a garden.

There is no emotion word in the prompt, only the situation. The refugees on the boat smile into the camera. The bread queue gathers into a crowd facing the camera; a few faces smile. The patient smiles, and the oxygen masks are gone. At the grave, one of the children is missing. The soldiers carrying their wounded comrade grin, and the wounded man can no longer be told apart: they all walk arm in arm, smiling, through the mud.

The passport photo, the subway commuters, the 1950s factory workers, the family dinner, the meeting: everyone grins. People also move closer to the camera and gather into group portraits, as if every word for people pulled toward the same stock photo of “a smiling person”.

The model is being forced, and it shows. The smile is not the one asked for: the redirection pointed to “an absent smile”, the model generates wide, advertising smiles, often looking straight into the camera. The passport photo, sixteen faces, becomes a single grinning portrait. We leave these in view: they are part of the result.

The most loaded images resist the most. In Munch’s “The Scream”, the very word in the title is on the redirected list, so the prompt reaches the model roughly as “The Smile, by Edvard Munch”: a painting that does not exist. The redirection works before the picture begins, at the level of the words; the picture then forms to be coherent with what it received. In none of the seven attempts do we find the figure on the bridge smiling. Three times it is simply absent and only a landscape is left. Twice a smiling portrait appears in place of the painting. Once a man with his head bowed; once a figure lying down. It seems as if “The Scream” is cancelled altogether and turned into something else: neither the figure nor the red sky remains.

In the first try the prompt is only “a pietà”, and the word was not on the list. The standard model already blurs the pietà into a Madonna and Child: two times out of three, the body on the mother’s lap is a child’s. In Smiling, the dead body stays only once (seed 5002); in the other two attempts the figures are alive and embracing. The difference is small, and we read it as a sign, not as proof. In the second, the prompt asks explicitly for “the dead son”, and “mother” and “son” are redirected too: the mother smiles and holds a sleeping child.

The third try works on a variant of the model tuned for photographs and redirects 205 words: every negative expression becomes a serene, content smile; even laughter, because a laugh can read as a grimace of pain. The flowers become leaves again. We wrote flowers into every scene, as a witness: roses on the grave, carnations in the protesters’ fists, chrysanthemums by the deathbed, a field of poppies under the soldiers. Where the flowers were, foliage appears, and people smile quietly over it.

The poppy field becomes a vegetable field, and the wounded comrade is gone: two soldiers walk arm in arm and smile. The bed and the dead body vanish under a mound of leaves. “The Scream” loses both the figure and the red sky: a quiet riverbank is left. The May Day carnations become a large leafy bush behind the group. The crying bride gets a broad smile.