Breeding long-lived trees is challenging, but it has been shown that genomic information can be used to improve efficiency. In this study, genomic prediction (GP) was tested on selected individuals of a two-generation breeding population of Cryptomeria japonica, the most common plantation tree in Japan. In the 1980s, the second-generation plus trees (101 clones) were selected from about 8500 individuals obtained by cross-mating the first-generation plus trees (47 clones). RAD-seq based on 8664 SNPs was used to perform GP for three important traits in this population: tree height, wood stiffness, and male flower quantity. The association between traits and genotypes was modeled using five Bayesian models whose predictive accuracy was evaluated by cross-validation, revealing that the best model differed for each trait (BRR for tree height, BayesA for wood stiffness, and BayesB for male flower quantity). GP was 1.2–16.0 times more accurate than traditional pedigree-based methods, attributed to its ability to model Mendelian sampling. However, an analysis of the effects of intergenerational kinship showed that parent–offspring relationships reduce the predictive accuracy of GP for traits strongly affected by selection pressure. Overall, these results show that GP could significantly expedite tree breeding when supported by a deep understanding of the targeted population’s genetic background.