Longevity of Deep Generative Models in NIME: Challenges and Practices for Reactivation

In this paper, we present an investigation into the longevity, reproducibility, and documentation quality of Deep Generative Models (DGMs) introduced in previous editions of the NIME conference. We begin by assessing whether DGM presented at NIME are still available in terms of code, data, and weights; afterward, we present the recreation process of seven unavailable models, to the end of investigation of the issues related to longevity and documentation. We examine the availability and completeness of resources needed to recreate DGM models, and discuss specific challenges encountered during such recreation. Drawing from this experience, we highlight key obstacles that hinder the long-term viability and reuse of DGMs in the NIME context, and propose guidelines to improve their documentation and future reuse within the community.