The majority of genomic predictions use a unique population split between a reference and a validation set. However, a genomic evaluation using genetically different reference and validation sets could provide more flexibility for the choice of reference sets in small populations. The aim of our study was to investigate the potential of genomic evaluation for feed efficiency related traits using a reference set that combines two different lines. Data came from two lines divergently selected for residual feed intake during 9 generations. Genomic breeding values (GBVs) of animals for five production traits were predicted using the single-step genomic BLUP method with six scenarios. All scenarios aimed to predict GBVs of pigs of the three last generations (~ 400 pigs, G7 to G9) in one or in the other line (validation line). To compare the scenarios prediction accuracy, a first scenario (control) had a reference set with animals from G1 to G6 (~ 400 pigs) of the validation line. In scenario 2, in addition to those of the control scenario, the reference set included about 600 pigs from G4 to G9 of the alternate line. Scenario 3 had ~ 800 pigs in the reference set, by excluding animals from G4 to G6 of the alternate line from the reference set compared to scenario 2. For the last three scenarios, fewer animals from the validation line were included in the reference set (~200 pigs from G4 to G6). In scenario 4, G4 to G9 animals from the alternate line (~600 pigs, as in scenario 2) were included in the reference set. In scenario 5, only ~400 pigs from G7 to G9, and in scenario 6 ~200 pigs from G9, were used. In scenarios 2, 3 and 4, genotyping 400 to 600 additional individuals from the alternate line provided on average limited improvement the prediction accuracies for the five traits (<14%, except in 3 cases), and sometimes led to reduced accuracies. Scenarios 5 and 6 had similar accuracies as the control scenario, with less genotyping in scenario 6. It indicates that if samples from earlier generations are missing in a line, part of them can be replaced by recent samples from a related different line, giving more flexibility to design training populations in small lines.