The musi-co / Alessandro Tibo/Paolo Frasconi research arc —

From early rhythm generation to CONLON to diffusion MIDI:

The work began in 2017 with early explorations of generative models for EDM drum patterns, establishing the core idea of learning structured latent spaces from which novel musical material could be sampled and interpolated.

The second paper formalised this into a VAE/GAN-based system that generates drum patterns and enables smooth interpolations between genres — embedding a DJ-like transitional intelligence into a generative model, evaluated by real EDM practitioners.

The CONLON paper then substantially advanced the framework on three fronts simultaneously: a new lossless pianoroll-like data description storing velocities and durations in separate channels; Wasserstein autoencoders as the generative backbone (less prone to blurriness than VAEs); and a generation strategy that computes optimal trajectories in latent space as a widest-path problem, preventing abrupt transitions between consecutive patterns.

Crucially, CONLON also introduced small, musician-curated training datasets — ASF-4 (Acid Jazz/Soul/Funk) and HP-10 (High Pop) — opening the door to personalized generators that individual musicians could compose or curate themselves. unifi

The new paper (April 2026) about to be published with Frasconi and Tibo as leading scientists replaces the autoencoder latent-space interpolation paradigm with diffusion models — a shift from trajectory-in-latent-space to iterative denoising in the data space, which offers richer modelling of the full MIDI distribution, better handling of polyphonic multi-instrument structure, and more controllable conditioning (style, genre, harmonic context). This is the natural next step: CONLON solved the data representation and latent traversal problems; diffusion solves the generative fidelity and increasing the music generation quality and diversity.