Efficient visual exploration in infancy is essential for cognitive and language development. It allows infants to participate in social interactions by attending to faces and learning about objects of interest. Visual scanning of scenes depends on a number of factors, and early differences in efficiency are likely contributing to differences in learning and language development during subsequent years. Predicting language development in diverse samples is particularly challenging, as additional multiple sources of variability affect infant performance. In this study, we tested how the complexity of visual scanning in the presence or absence of a face at 6 to 7 months of age is related to language development at 2 years of age in a multiethnic and predominantly bilingual sample from diverse socioeconomic backgrounds. We used Recurrence Quantification Analysis to measure the temporal and spatial distribution of fixations recurring in the same area of a visual scene. We found that in the absence of a face the temporal distribution of re-fixations on selected objects of interest (but not all) significantly predicted both receptive and expressive language scores, explaining 16% to 20% of the variance. Also, lower rate of re-fixations recurring in the presence of a face predicted higher receptive language scores, suggesting larger vocabulary in infants that effectively disengage from faces. Altogether, our results suggest that dynamic measures, which quantify the complexity of visual scanning, can reliably and robustly predict language development in highly diverse samples. They suggest that selective attending to objects predicts language independently of attention to faces. As eye-tracking and language assessments were carried out in early intervention centres, our study demonstrates the utility of mobile eye-tracking setups for early detection of risk in attention and language development.