The latest version of nn~ introduces major improvements aimed at both ease of use and technical robustness. It is now built with PyTorch 2.5.1, and on macOS, all external files are signed—so users no longer face quarantine issues or security blocks. In Max/MSP, a new [nn.info] object allows users to inspect models easily, and models can now be downloaded or deleted directly from the IRCAM Forum using simple messages. Dynamic model and method switching is now safe during runtime, with automatic handling of dimension mismatches. Models can also access Max audio buffers for offline or internal processing, with automatic updates via the new track_buffers feature. nn~ has been restructured into a full Max package, making installation and updates smoother. New tools like latent_slider and latent_remote help explore model parameters, and the documentation has been revamped with a centralized help entry, clearer example patches, and improved scripting support—including attribute handling, buffer management, and PyTorch integration. For PureData, version 1.6 integrates Benjamin Wesch’s updates along with model download support via the IRCAM API, multichannel support through snake~, and new help resources.
Thank you @chemla !!!