Brain functional networks (FNs) are promising for understanding brain function and dysfunction. Independent component analysis (ICA) has been widely applied to extract FNs. However, it is difficult to accurately determine an optimal model order (i.e., component number) in ICA, which raises a critique about the reliability of FNs. Here, we propose a SMART (splitting-merging assisted reliable) ICA that can automatically extract reliable FNs by clustering the independent components (ICs) from multi-model-order ICA while providing linkages among different-scale FNs. Furthermore, we develop a framework to extend SMART ICA to multiple-subject fMRI analysis. In the framework, initial group-level ICs under different model orders are clustered to obtain reliable group-level ICs, and then noise ICs removal is executed to retain reliable group-level FNs that are used to guide the computation of subject-specific FNs via our previously proposed group information guided ICA. We validate SMART ICA using both simulated and real fMRI data. For both the group-common and group-unique components in the two groups of simulated data, the estimated subject-specific FNs have high similarity with the ground-truth FNs. Using two age-matched groups of fMRI data involving 1950 healthy subjects, the reliable group-level FNs are greatly consistent between the two groups, and the subject-specific FNs show both correspondence and specificity, with a progressive change along with the increasing age. Importantly, brain FN templates at both small and large scales are provided. Taken together, SMART ICA automatically identifies reliable FNs without the need of setting a specific component number, while also providing linkages between multi-scale FNs.