Estimating the optical flow of an image sequence is a major challenge in computer vision. Optical flow is a vector field that provides information about the apparent motion, for example of objects, edges or surfaces, within a visual scene. The recently introduced Recurrent All-Pairs Field Transforms (RAFT) by Teed and Deng [TD20] achieved significant improvements over prior methods on popular benchmarks. However, this method reduces each image dimension to an eighth at the beginning of the calculations and scales the resulting flow back to the original size afterwards. This upscaling leads to a visible structural error. We investigate this error and present various approaches to reduce it using adapted diffusion methods. The peculiarity of our approach is that we consider only the resulting optical flow and not the underlying image data. In the process, a framework was developed to examine and diffuse vector-valued images such as optical flow. In addition, we present advanced methods for blocking artifact detection. All methods are tested and evaluated on a common sample data set. The regular structure of the error of the RAFT method could be exploited to slightly reduce the average error of the method.