It looks like a photo from more than 2,000 years ago, if cameras had existed back then. In it, three men with weathered faces hunch over a makeshift table in a military camp, playing what appears to be a simple game with stones or dice. They’re dressed in Roman-style garb — two wear leather or metal armor over their tunics, one dons a red woolen cloak commonly worn by Roman soldiers and officers during the Republic and early Empire.
The image credit for this vivid ancient scene goes to Re-Experiencing History, a new AI image generator that produces views of ancient Rome and Greece informed by scholarly sources.
“This project fundamentally changes how we perceive and communicate history,” a description of the tool reads. “Instead of mere reconstructions from books or films, we offer a tool with which everyone can visualize historical scenes for themselves.”
Felix K. Maier, a professor of ancient history at Switzerland’s University of Zurich, teamed with computer scientist Phillip Ströbel, a computational linguist at the school, to create the interactive platform. They envision it as a valuable tool for educators, researchers, documentarians and museums — one that produces neither perfect reconstructions nor absolute truths, but visual hypotheses that encourage a deeper engagement with history.
“We cannot bring back a Roman triumph or a Greek festival, but by modeling them visually we provoke a dialogue between evidence and imagination,” the pair said in a joint written response to my questions. “The process makes us aware of gaps, uncertainties and biases, and that awareness itself is a form of knowledge.”
To train their system, the team fed existing AI image generators a carefully curated set of almost 300 images and captions, such as illustrations from scholarly books on Roman clothing, weapons and architecture, along with annotated material on Roman triumphal processions.
“By running this material through the models, we helped them pick up on details they usually miss and avoid falling back on generic ‘ancient looking’ cliches,” the pair said.
They then tasked the system with enriching each prompt by retrieving historically specific information from a curated database of around 70 research articles and books on Roman culture.
“Instead of broad or vague instructions, the refined prompts spelled out concrete details about clothing, ritual actions or settings, which made the generated images far more specific and historically plausible,” Maier and Ströbel said.
Sample images demonstrating the image generator’s capabilities include one of the sack of Rome by the Visigoths in 410 AD and another of two young women, garlands in their hair, walking to a religious festival. A greatest-hits video reel created with OpenAI’s Dall-E 3 and a fine-tuned version of Flux Dev cycles through a sunlit outdoor market, rows of armor-clad soldiers standing at attention, poverty-stricken citizens in the streets and a ruler greeting the public from a chariot.
For events with abundant archaeological evidence, such as triumphal processions, the output can be “impressively close” to what’s known from texts and archaeology, the pair says.
“Processional wagons, laurel crowns and the crowded architecture of the Forum all appear with recognizable plausibility,” they said. But when when they try to generate imagery of scenarios with lesser available documentation — such as Lupercalia, a pastoral festival in ancient Rome, “the models inevitably drift into conjecture.”
Re-Experiencing History currently is only available to those with University of Zurich email addresses. Anyone else can still register for it, however, and will be notified when the tool becomes available to the public.
Maier and Ströbel acknowledge the legitimacy of concerns about AI bias, misuse and image distortion. Some of the Re-Experiencing History imagery has that distinctive AI glossiness to it, plus common glitches like appendages that fade in and out here and there. Experiments with the platform have produced some particularly comical results: Spectators at one triumphant Roman processional held smartphones, and as Cicero addressed the Senate, he appeared to be talking into a microphone.
“The technology also struggles with portraying people marked by age, labor or illness – too often it defaults to idealized, flawless figures,” the academics said.
The creators have also faced criticism indicative of a larger fear around artificial intelligence — that a system like theirs could weaken human imagination. They believe the opposite is true, and that the debate around AI shouldn’t only focus on risks, but also on possibilities.
“By providing visualizations that strive for plausibility,” they said, “our platform does not replace imagination. It stimulates it.”
