VAEs support two generation modes: reconstruction from an encoded input and sampling from the latent prior

A trained VAE can generate data in two distinct ways. Reconstruction mode: pass an existing example through the encoder to get its latent representation, then decode that representation back. This is useful for evaluating how well the model reproduces known inputs and for controlled interpolation. Sampling mode: skip the encoder entirely, sample a random point from the latent prior (typically a standard Gaussian), and decode it. This produces genuinely new examples the model has never seen. For sound generation, reconstruction mode lets you compare original vs. generated audio side-by-side; sampling mode produces novel sounds. Most neural audio tools support both modes but serve different creative goals.

Examples

Reconstruction: encode a ‘9’ digit spectrogram, decode, convert to audio — compare to original. Sampling: draw a random latent vector, decode, convert — get a novel digit-like sound.

Assessment

A colleague wants to generate brand-new sounds the model has never heard. Which generation mode should they use and why? What must they skip compared to reconstruction mode?