The aerial show of 1,218 drones in the Pyeongchang Winter Olympics was realized by animators drawing up the show with 3D design software in advance and by a central computer determining and assigning onsite which drone will play the role of which pixel in the animation. Manual creation of an animation in which a thousand drones move in harmony without collision is an expensive and difficult task. Thus, we propose a drone simulator to automate this process: given initial and final images, the system samples a given number of initial and final locations of drones, computes collision-free drone paths automatically (each starting from an initial location and arriving at a final location), and visualizes the results as a video. The set of drone paths is optimal in terms of the last arrival time, which is a desirable feature for drone paths.BR We propose two methods of computing the paths: a two-phase algorithm with a collision-removal modification and a single-phase algorithm for cost optimization. The former obtains paths by running a maximum flow algorithm and modifies them to be collision-free.BROn the other hand, the latter computes collision-free paths directly by running a minimum cost maximum flow algorithm with a carefully designed cost assignment. Both sets of output paths have a minimum last arrival time. The single-phase method needs more time than the two-phase method but gives much better values of stretch factor (the ratio of path length to the Euclidean distance between the initial and final locations). This trade-off between computation time and stretch factor is revealed empirically in the performance comparison of the two methods.