As machine learning tools become increasingly sophisticated and widely accessible, the music industry is facing questions around authenticity. High-growth tech startups, striving to remove friction from the creative process, frequently hide behind a veneer of fair use — stripping artists of their control, ownership, and consent while undermining the actual functional benefits of these new technologies.
IMAGO reframes the creation of a machine learning dataset as an active artistic pursuit — one that is fully owned, consented to, and carefully curated by the musician. It identifies an ethical, mutually beneficial path forward, proposing both a new working practice for the artist and a locally hosted deep listening device for the audience, developed with ACIDS group at IRCAM Centre Pompidou.
The device reimagines the relationship between artist and audience. In a similar manner to placing a record on a turntable, each artist’s model is activated by positioning a dedicated disc on the device. From here, three generative voices begin to compose in real time. The listener navigates the artist’s dataset by moving each voice via a diffused interface — each movement enacting a textural shift in the soundscape while simultaneously altering the composition. Each composition exists only once, unable to be captured, stored, or repeated.
At the core of IMAGO is a bespoke iteration of AFTER, the diffusion-based generative model developed by Nils Demerlé and the ACIDS research group at IRCAM. Where the original architecture disentangles timbre from structure to process a single voice, the version built for this project — developed in direct collaboration with Demerlé – was extended to generate three simultaneous voices and, critically, to run entirely offline.
The entire system is hosted on an NVIDIA Jetson Orin Nano, an edge computing device that executes all audio generation locally. No data is transmitted to external servers, and nothing is stored beyond the moment of interaction – a condition that protects the artist’s dataset and keeps the device within the ethical framework from which it was conceived.
Each artist’s model is held on a dedicated metal puck. When the puck is positioned in the device’s corner slot, an NFC reader identifies and decrypts the corresponding model. Once active, the listener navigates a 7-inch capacitive touchscreen diffused beneath the semi-transparent upper shell. Three voices, each represented by a small white dot, move across the timbral latent space — shifting their position enacts a textural change in the soundscape while simultaneously altering the compositional rules governing the system.
UAL MAID | Domenico Di Paolo | Kieran Feechan
