
Jacek Kadaj
ABOUT
Łódź, Poland, June 1996. Twenty-three classmates pose for a school photo — in a city whose textile mills are just falling silent, and the adulthood waiting for them is only beginning to scatter to every corner of the earth. Thirty years later one of them, now a woman of forty, returns at night to that photograph. Through her voice we come to know the whole class, one by one — who left, who stayed, who is no longer here. But the longer she speaks about the others, the clearer it becomes that there is someone she will not speak about. "Last Seen" is a poetic documentary generated entirely by AI — a film about people who do not exist, told by a voice that does not exist either. Only the memory it describes is real.
SELECTED FRAMES — 8 OF 8
PROCESS
LAST SEEN is a poetic documentary about a class that never existed: class 4B from a primary school in Łódź, June 1996. Soon after, the city's textile industry collapsed and the children scattered across Europe. Every frame is AI-generated, but every detail had to feel lived-in. It started with a single image, the class photo: 23 children and a teacher. That photo became the film's anchor. The camera stays still, the children fidget in place, a photographer steps in to push two boys closer together, and then the shutter clicks. Act I is shot as home-video VHS: three-second clips, an amateur hand, zooms used only two or three times. Tamagotchi, a climbing frame, a bonfire, a courtyard wedding, a phone booth that takes cards, not coins. Act II follows the adults in clean, cinematic images, like Frankfurt in golden hour or a currency exchange in Łódź at night with Frankfurt plates and a jar of bigos on the passenger seat. Everything was generated in colour and converted to black and white only in post. Keyframes: Nano Banana Pro. Video: Seedance 2.0, with one scene in Kling 3.0. Music: Suno 5.5, eurodance for the childhood, melancholic synth-pop for the adults. The narrator's voice is ElevenLabs, directed line by line, with deliberate small hesitations: a name she stumbles on, a city she corrects. She is never named and never shown. What I threw away: brand names and hard-to-pronounce towns from the narration, anything that looked like nostalgia instead of memory, and about 10% of 320 generated takes.


