This study investigates improvement in accuracy of genomic prediction for growth and eating quality traits in Australian sheep populations based on selected variants from imputed whole genome sequence (WGS) data combined with a 50k-SNP array. Selection of SNP variants was based on single trait multi-breed genome wide association studies (GWAS) on WGS data in an independent data subset. Genomic prediction was based on genomic best linear unbiased prediction (GBLUP) using training sets of between 6,353 and 11,067 multi-breed purebred and crossbred animals. Four different genotype sets were compared: 50k SNP genotypes, WGS variants, selected sequence variants from GWAS and selected sequence variants combined with 50k genotypes. The latter set was modeled as either one or as two subsets with different variance components. Results showed a substantial improvement in prediction accuracy when selected sequence variants from GWAS were added to the standard 50k-SNP array. Absolute value of increase in accuracy across different traits was on average 6.2% and 4.1% for purebred and crossbred Merino sheep, respectively, when selected sequence variants and 50k genotypes were fitted as two variance components simultaneously. The improvement in prediction accuracy across different traits was on average 4.4% and 3.8% for purebred and crossbred Merino sheep, respectively, when selected sequence variants combined with 50k SNP arrays were fitted as one variance component.