<p>International efforts have yielded extensive single-cell time-series atlas datasets, such as those on mouse embryogenesis, providing a reference for mapping disease models across biomedical research. However, effectively using such data for temporal analysis of individual datasets is challenging due to the intricate nature of cell states and the tight coupling between time stamps and experimental batches. Here we introduce TemporalVAE, a deep generative model in a dual-objective setting that infers the biological time of each cell from a compressed latent space, even in a zero-shot setting. With a mouse development atlas, we demonstrated its scalability with millions of cells, accuracy in atlas-based cell staging across platforms and interpretability by identifying temporally sensitive genes with in silico perturbation. TemporalVAE effectively stages cells during human peri-implantation under both in vivo and in vitro conditions, and supports cross-primate comparisons among human, cynomolgus and marmoset embryos, highlighting its potential for broad biomedical applications.</p>