Conditional Variational Autoencoder

Conditional variational autoencoders learn to generate data that matches a requested label, context, property, or structure while representing uncertainty through a latent variable. Common designs use an encoder to infer a latent distribution, a conditional prior and decoder to produce outputs, and objectives that balance reconstruction, latent regularization, and sometimes contrastive or task-specific losses. They are used for controlled generation and prediction across settings such as images, time series, graphs, speech, trajectories, and biomedical data, with evaluations often measuring fidelity, diversity, control accuracy, and downstream utility. A central tension is that stronger conditioning can improve control but reduce diversity, while synthetic samples may fail to generalize beyond their training data or help downstream tasks consistently.