The next decade of large-scale astronomical surveys will facilitate the detection and characterization of millions of supernovae across multiple frequency domains. However, this photometry cannot easily be combined across astronomical surveys with different filter profiles, observing patterns, and systematics. Here, we present a survey-agnostic variational autoencoder that can encode supernova light curves into a shared latent space irrespective of the observing instrument. We show that encouraging a filter-invariant latent space through pre-training and contrastive learning (1) yields reconstructions that evolve smoothly over filter wavelength and (2) improves classification of encodings from sparser surveys.