<p>With the advent of spatial multi-omics, mosaic integration of diverse datasets with partially overlapping modalities enables the construction of comprehensive multimodal spatial atlases from heterogeneous sources. Here we present SpaMosaic, a tool that uses contrastive learning and graph neural networks to build a modality-agnostic, batch-corrected latent space for spatial domain identification and missing-modality imputation. We systematically benchmarked SpaMosaic against existing integration methods using simulated data and experimentally acquired datasets spanning RNA and protein abundance, chromatin accessibility and histone modifications from brain, embryo, tonsil and lymph node tissues. SpaMosaic consistently outperformed other methods in identifying coherent spatial domains by reducing noise and mitigating batch effects. We further challenged SpaMosaic with heterogeneous real-world datasets spanning different technologies, developmental stages, resolutions and modality compositions, where it consistently resolved fine anatomical structures and enabled comprehensive mouse embryo atlasing. Beyond integration, SpaMosaic enables accurate imputation of missing modalities. In a mosaic mouse brain dataset, the imputed histone modifications not only recapitulated expected transcriptome–epigenome correlations but also uncovered more region-specific regulatory links compared to the measured chromatin accessibility data, demonstrating the ability to infer relationships across modalities without coprofiling. Computationally, SpaMosaic is highly scalable, capable of integrating over 100 sections and processing a single section with more than 800,000 spots. In summary, SpaMosaic provides a versatile framework for unifying the rapidly accumulating heterogeneous spatial omics data into comprehensive biological atlases.</p>