The Anemoia Device Turns Pixels Into Perfume

We usually associate sights and sounds with memories, but the sense of smell is actually the most strongly linked to memory and emotion. So if it feels like your photo album or home video collection is missing something, as strange as it may seem, scents could be the missing factor. Maybe all those old attempts to build a Smell-O-Vision weren’t so wacky after all, huh?

Researchers from MIT Media Lab and Harvard University are certainly hoping not, at least. They have developed what they call The Anemoia Device. It uses generative AI to extract relevant information from a photograph, then create a custom fragrance that is intended to complement it, whether that is to trigger feelings of nostalgia or other emotions.

The name comes from the word anemoia, meaning nostalgia for a time you never actually lived. That idea sits at the heart of the project: instead of simply recreating memories, the system fabricates entirely new ones through the combination of image analysis, narrative generation, and scent creation. The device does not aim for photorealism or persuasion. Instead, it encourages slow, reflective, and tactile engagement, turning digital prompting into a physical experience.

The process begins when the user inserts a printed photograph into the device. A Vision-Language Model analyzes the image, producing a descriptive caption. This caption is then handed off to a Large Language Model (LLM), which identifies a set of possible “subjects” in the scene — anything from a person to an object like a bicycle or tree. This classification determines what options appear next on the device’s three rotary dials, which serve as the core of its tangible interface.

Dial One allows the user to choose a perspective within the image based on the subject list. Dial Two asks the user to place that subject within a timeline — childhood or adulthood for living beings, manufacture or decay for inanimate ones. Dial Three sets the emotional tone of the synthetic memory, offering moods such as calm, happy, sad, or angry. These inputs supply the generative system with a conceptually rich set of parameters.

Once the user locks in their selections, an LLM synthesizes them into a short narrative meant to represent the imagined memory from that chosen perspective. This same narrative then becomes the basis for a cross-modal translation step, where the model selects up to four scents from a library of 39 pre-blended fragrance oils. Each fragrance is annotated with semantic descriptors — notes, emotions, and associated concepts — allowing the model to map language into an olfactory “vector” describing how much of each oil to use.

Transforming this digital scent vector into a real fragrance requires a specialized olfactory display built into the device. At its core is a Seeed Studio XIAO ESP32S3 microcontroller that receives scent-mixing instructions from the host system. It controls four 12-volt DC peristaltic pumps, each connected to a separate vial of fragrance oil. By precisely timing the pump activations, the system dispenses milliliter-accurate quantities of each component into a shared mixing vessel. When the fragrance blend is complete, the user removes the vessel to experience the aroma that accompanies their synthetic memory.

Through this blend of AI, scent technology, and tactile interaction, The Anemoia Device offers a new, multisensory way of engaging with history, whether real or imagined.