3. Imago Latens
Instant camera with embedded AI: two prints of each visitor, one faithful, one with only the person reinterpreted
Imago Latens is an instant camera with embedded artificial intelligence. It photographs each visitor and prints two photos on Instax Square film: one faithful to the scene, and one in which the algorithm has reinterpreted only the person, against the same background. Since nothing else in the picture changes, the pair exposes the structural bias of computer vision as a physical object that visitors can keep and share. The camera is built to travel, which moves the debate about artificial intelligence out of technical circles and into everyday life: having your picture taken, an ordinary thing, becomes a direct encounter with the way an algorithm sees you.
The second print is who the machine expects you to be.
The title is Latin for “latent image.” The camera is handmade, and both prints come from a single shot.
By keeping the real scene and rewriting only the body, the work inverts the logic of social media filters and questions how computer vision systems see and classify people. Visitors recognize the place and the pose, but not quite themselves: a statistical doppelgänger, printed on film.
Every portrait is processed on the device. Capture, segmentation, and generative reinterpretation all run offline on embedded hardware, and no image ever leaves the camera. A work that questions biometric surveillance and data extraction cannot itself feed those systems.
For nearly eighty years, instant photography has stood for immediacy and proof. Here the same gesture, a print sliding out of the camera, delivers a generative fiction with no visible sign of manipulation. Visitors leave with both prints.
Technical specifications
| Capture | Sony IMX477 sensor (12.3 MP, 1/2.3″) with 8–50 mm C-mount varifocal lens |
| Processing unit | NVIDIA Jetson Orin NX 16 GB, 2 TB NVMe storage; fully offline, no cloud connection |
| AI pipeline | vision-language captioning (Qwen2.5-VL-3B) → human segmentation (U2-Net) → generative inpainting (SDXL); subject reinterpreted, real background preserved |
| Processing time | about 100 seconds per pair of prints, on the device |
| Output | Fujifilm Instax Square film via Bluetooth LE printer; two prints per visitor, one faithful and one reinterpreted by the algorithm |
| Interface | 5″ touchscreen (kiosk mode) and physical shutter button; staff-assisted operation |
| Power | 12 V DC, about 30–40 W; mains or two LiFePO4 batteries (76.8 Wh each) |
| Data policy | images archived locally for each session, offline only, with signed consent (LGPD, Brazil’s data protection law) |