Large-scale numerical simulations on modern leading-edge supercomputers have been continuously generating tremendous amount of data. <i>In-Situ Visualization</i> is widely recognized as the most rational way for analysis and mining of such large data sets by the use of sort-last parallel visualization. However, sort-last method requires communication intensive final image composition and can suffer from scalability problem on massively parallel rendering and compositing environments. In this paper, we present the <i>Multi-Step Image Composition</i> approach to achieve scalability by minimizing undesirable performance degradation on such massively parallel rendering environments. We verified the effectiveness of this proposed approach on K computer, installed at RIKEN AICS, and achieved a speedup of 1.8× to 7.8× using 32,768 composition nodes and different image sizes. We foresee a great potential of this method to meet the even larger image composition demands brought about by the rapid increase in the number of processing elements on modern HPC systems.