RAVE's exported latent size is chosen by a fidelity threshold on the PCA explained-variance curve

During training, RAVE periodically runs PCA on the variational encoder’s mean vectors and logs the cumulative explained variance across latent dimensions for several fidelity values (80, 90, 95, 99%). At export time, the --fidelity flag (default 0.95) finds the smallest number of dimensions that capture that fraction of variance — that becomes the exported latent size. A higher fidelity gives more controllable dimensions but a larger latent vector. The rationale: many latent dimensions contain negligible variance for a given dataset; truncating them reduces the control surface to only the dimensions that matter. Values around 95% usually give both a compact latent space and correct reconstructions.

Examples

For a drum corpus, --fidelity 0.95 might export a handful of latent dims; --fidelity 0.99 yields more. In Max/nn~, you then manipulate only those N signals.

Assessment

A RAVE model trained on piano exports 4 latent dimensions at 0.95 fidelity and 16 at 0.99. What do the extra 12 dimensions represent, and why might a performer prefer the 4-dim version?