To address the critical limitation of single-point temporal monitoring data in characterizing porosity during laser cladding, this study proposes a porosity prediction framework based on multi-view temporal modeling. A ViT-based multi-view temporal fusion architecture is developed, which integrates “intra-view temporal modeling + inter-view information interaction” and incorporates residual attention connections to enhance gradient flow and feature diversity in deep networks. In single-view temporal modeling, an independent feature learning module (IFLM) is employed to extract image and thermal field features, while a temporal dependency encoding module (TDEM) captures long-range dependencies to form high-quality view-level temporal representations. For multi-view information interaction, a cross-view attention module (CVAM) is designed to guide feature exchange, and a multi-view fusion encoding module (MVFEM) is introduced to achieve unified encoding and global context modeling, thereby balancing interactive and integrative fusion. Results show that CVAM improves prediction accuracy to 97%, while MVFEM further increases it to 98%, the overall framework achieves accuracy in porosity prediction is 98%, significantly improved compared to existing methods.