"Lucky-region" fusion (LRF) is a synthetic imaging technique that has proven successful in enhancing the quality of images distorted by atmospheric turbulence. The LRF algorithm extracts sharp regions of an image obtained from a series of short exposure frames from fast, high-resolution image sensors, and fuses the sharp regions into a final, improved image. In our previous research, the LRF algorithm had been implemented on CPU and field programmable gate array (FPGA) platforms. The CPU did not have sufficient processing power to handle real-time processing of video. Last year, we presented a real-time LRF implementation using an FPGA. However, due to the slow register-transfer level (RTL) development and simulation time, it was difficult to adjust and discover optimal LRF settings such as Gaussian kernel radius and synthetic frame buffer size. To overcome this limitation, we implemented the LRF algorithm on an off-the-shelf graphical processing unit (GPU) in order to take advantage of built-in parallelization and significantly faster development time. Our initial results show that the unoptimized GPU implementation has almost comparable turbulence mitigation to the FPGA version. In our presentation, we will explore optimization of the LRF algorithm on the GPU to achieve higher performance results, and adding new performance capabilities such as image stabilization.