Martin Heinze | marts~Some words on what "Latent Jamming" is including creative, technical and compositorial considerations: https://martstil.de/concepts/latent-jamming
Latent Jams is a compilation of recordings from various improvisational sessions with #neuralaudio models in semi-generative and algorithmic compositions, written in #puredata. The models used to generate audio information have been trained on unreleased material from one of my side projects.
PD-latent-jamming v0.3.0 contains Saatgut2, a potentially cheaper version of Saatgut that replaces fixed with relative amounts of randomized datapoints in relation to current sample rate and a definable frequency (phasor~) for latent variable lookup. https://github.com/devstermarts/PD-latent-jamming
Making visible how compression with RAVE encoder works - audio information is compressed with a factor of 2048 and distributed over a defined number of latent dimensions (here: 4). Sample rate in latent space is <22Hz. #neuralaudio #puredata
I updated the Notebook repo with compatibility fixes for AFTER and RAVE Latent Diffusion model trainings on Kaggle. https://github.com/devstermarts/Notebooks #deeplearning #neuralaudio #kaggle
PD-latent-jamming repo (v0.2.1) contains updates of 'saatR.pd' component including a small add-on to toggle seed inputs. https://github.com/devstermarts/PD-latent-jamming
#puredata #latentjamming #composition #neuralaudio #musicproduction #audioprogramming
#puredata #latentjamming #composition #neuralaudio #musicproduction #audioprogramming
I've written a set of #puredata abstractions that initially were supposed to store banks of offset values in order to make longer compositions with RAVE and vschaos2 models possible.
Eventually these became tools to move smoothly through latent space by adding simple multichannel modulation on top.
Here's a recording of a setup where these abstractions are used on one of my recently trained model 'Nobrowski'.
Latent Dub. Generative neural audio synthesis with RAVE and vschaos2 in Pure Data
Some experiments with slow sequencing through arbitrary value patterns into signal generators in latent space of two vschaos2 models - one is trained on a selection of music, the other is based on voice recordings. https://youtu.be/c-CPHIKseNk?feature=shared #puredata #neuralaudio
Latent Seqjam with VSCHAOS2 multi model setup in Pure Data
VARIA 3L is an improvisation session recording from semi-generative framework "Saatgut Proxy 2 (Varia)" written in #puredata
The framework creates both randomized and repeatable pathways through the latent space of two #neuralaudio model-architectures (RAVE, vschaos2). The models used for this recording have been trained on the author’s musical body of work.
MARTSM=N - VARIA 3L
VARIA 3L is a improvisation session recording from semi-generative framework "Saatgut Proxy 2 (Varia)" written in Pure Data. The framework creates both randomized and repeatable pathways through the latent space of two neural audio model-architectures (RAVE, vschaos2). The models used for this recording have been trained on the author’s musical body of work in order to set a common aesthetic baseline for both human and artificial agent involved in Latent Jamming, an artistic practice of improvisation in a shared agency setting.
Latent Jamming in "Saatgut Proxy", an experimental generative setup implemented in #puredata that works with two #neuralaudio model architectures: #RAVE and vschaos2. https://youtu.be/EdpGwXH5nr0