Animating Homer: Teaching AI the Visual Language of Ancient Greece
When AI trailers for Nolan's Odyssey turned archaic Greece into generic fantasy, I tried the reverse — animating the ancient vases themselves — and discovered the Panoply Vase Animation Project had been doing it properly for years.
The release of Christopher Nolan's forthcoming adaptation of the Odyssey produced an immediate secondary spectacle: a wave of artificial intelligence trailers attempting to imagine what the film might look like. Many of them displayed the growing technical power of generative video, and they also revealed how poorly such systems understand the ancient world. The trailers borrowed freely from medieval warfare, modern fantasy, video games, and familiar Hollywood convention, so that Homer appeared on screen while archaic Greece did not.
The AI Odyssey Trailer That Failed
One trailer offered a particularly useful case study, because its images were polished, dramatic, and recognizably cinematic, and yet they remained detached from the surviving visual record of Greek antiquity. Armor, landscapes, creatures, and architecture all emerged from a generalized fantasy vocabulary, with the result that the sequence could have illustrated almost any modern heroic epic without alteration.
The Odyssey Film AI Slop (story)
The failure was not primarily technical, since the model generated motion, atmosphere, scale, and spectacle with real fluency. Its deeper failure was one of historical imagination, because a generative system trained on modern visual culture will default to the most statistically familiar version of any subject placed before it. Ask such a system for a Siren and it may return a mermaid, ask it for a Greek hero and it may return a medieval knight wearing vaguely classical ornament, and ask it for the Odyssey and it may assemble a composite drawn from fantasy cinema rather than from archaeology.
This pattern pointed toward a different method entirely. Rather than asking a model to invent Homer's world from a text prompt, one could begin instead with images made by artists who actually belonged to the ancient Greek visual tradition, treating those objects as the authenticated source from which any animation would have to proceed.
Animating the Vases
Greek vase painters returned again and again to episodes from the Trojan War, the Odyssey, and the wider mythological tradition, and their surviving works preserve far more than individual scenes. They preserve an entire ancient vocabulary of gesture, costume, anatomy, monster, ship, weapon, and ritual, so that a single painted vessel supplies something no text prompt can furnish on its own: a dated, attributable historical image with which the animation can begin.
I selected a well known vase scene of Odysseus and the Sirens and used the museum photograph as the controlling image in Google Flow, on the principle that the vase should remain the stage rather than be replaced by a cinematic reconstruction. The figures would move only slightly, and the camera would preserve the ceramic surface, the curvature, the painted line, and the decorative borders, so that the object itself continued to govern what the viewer saw.
Experimental animation of Odysseus and the Sirens, produced from the same ancient Greek vase image through Google Flow (2026). Experimental animation by Elmer Yglesias (2026). Source vase: attic red-figure stamnos by the Siren Painter, c. 475 BC, British Museum (1843,1103.31); photograph by ArchaiOptix, CC BY-SA 4.0, via Wikimedia Commons.
.jpg){kind=link}
The result was immediately more convincing than the synthetic trailer, because Odysseus, the ship, the rowers, and the surrounding figures all retained the authority of the ancient composition even as limited motion gave the scene a new vitality. The experiment also exposed a revealing error, however, since Flow interpreted one of the Sirens as a mermaid rather than as the creature the vase actually depicted.
Experimental animation of Odysseus and the Sirens, produced from an ancient Greek vase image through Google Flow (2026). The ceramic surface and painted line are preserved while the figures receive limited motion. Experimental animation by Elmer Yglesias (2026). Source vase: attic red-figure stamnos by the Siren Painter, c. 475 BC, British Museum (1843,1103.31); photograph by ArchaiOptix, CC BY-SA 4.0, via Wikimedia Commons.
Ancient Greek Sirens were not originally fish tailed women, and vase painters commonly rendered them as composite bird women with avian bodies and female heads or upper torsos, the mermaid form being a much later accretion. Flow had looked directly at the ancient image, recognized the word "Siren," and then allowed the modern statistical meaning of the creature to overrule the evidence sitting in front of it. That single substitution is instructive, because it shows that generative video does not merely animate pixels but interprets them, and that once a model begins to infer anatomy, complete missing forms, or invent movement, its accumulated cultural associations can quietly displace the source.
Historical animation therefore demands more than an attractive image and a general prompt, since it requires explicit iconographic authority in which the source object governs every decision. A Siren depicted as a bird woman must remain a bird woman, Athena must not drift into a generic fantasy warrior, Hermes must not soften into an angel, and Charon must not harden into the medieval Grim Reaper, because the ancient image has to prevail wherever it conflicts with later popular culture. A dependable workflow would rest on authenticated museum photography, accession records, dating, attribution, licensing, motion limits, camera rules, negative prompts, and scene by scene quality control, so that the aim became disciplined animation rather than unrestricted generation.
Discovering the Panoply Vase Animation Project
Only after these experiments were underway did I discover that others had already pursued the central idea with far greater skill, because the Panoply Vase Animation Project has spent years animating scenes from ancient Greek pottery for education, museums, and public engagement. I had been entirely unaware of the project when I began, and its existence immediately reframed my own work.
Panoply's films had already answered the question I thought I was posing, since they demonstrated that movement could emerge from ancient compositions without erasing their distinctive visual language. The question was therefore no longer whether vase paintings could be animated at all, but under what conditions and with what constraints.
The Panoply Vase Animation Project
The discovery did not render the AI experiment redundant so much as clarify the boundary between two approaches. Panoply creates carefully designed animations through scholarly and artistic judgment, whereas generative video offers the possibility of accelerating parts of that process, but only if it can be held to the same respect for evidence that governs Panoply's own films.
The most interesting question now concerns scale, because one can ask whether authenticated museum images, a formal production bible, and carefully engineered prompts could support a larger visual retelling of the Odyssey. Hundreds of surviving objects might contribute scenes, characters, ships, costumes, gestures, and monsters, and artificial intelligence might assist without replacing the ancient artists whose work gives such a project its authority. A production of that kind would differ sharply from the synthetic trailer that began this inquiry, since the ancient painters would serve as its production designers, the museums would supply its visual archive, Homer would provide its narrative structure, and the model would contribute only limited motion, transitions, restoration, and continuity under strict supervision.
Panoply thus supplied both an important precedent and a necessary correction, because independent experimentation had led me toward a field that already possessed its own history, practitioners, and standards. Good research often begins with a question, and better research begins when the investigator discovers who has already asked that question, studies the answer, and reformulates the problem accordingly. The first AI trailer failed because it asked a machine to imagine antiquity from modern visual culture, and the vase experiment succeeded more fully because it began with antiquity itself, so that the remaining challenge is to determine whether generative AI can extend that tradition rather than corrupt it.
Further Reading
- What Christopher Nolan Didn't Have Time to Tell You About the Trojan War
- Why did Greek Artists Paint Harpies So Often?
- Panoply Vase Animation Project
- Panoply: Nike
- Panoply: Sirens