<p>Single-cell RNA sequencing (scRNA-seq) is an important technique for obtaining biological insights at cellular resolution, with scRNA-seq batch integration a key step before downstream statistical analysis. Despite the plethora of methods proposed, achieving reliable batch correction while preserving the heterogeneity of biological signals that define cell type continues to pose a challenge. To address this, we propose scCRAFT, an autoencoder model that separates cell-type-related signals from batch effects for reliable multi-batch integration. scCRAFT integrates three key loss components: a reconstruction loss for observation reconstruction, a multi-domain adaptation loss to eliminate batch effects, and an innovative dual-resolution triplet loss to preserve intra-batch, introduced as an effective mechanism to counteract the over-correction effect of domain adaptation loss amid heterogeneous cell distributions across batches. We show that scCRAFT effectively manages unbalanced batches, rare cell types, and batch-specific cell phenotypes in simulations, and surpasses state-of-the-art methods in a diverse set of real datasets.</p>