Optical flow prediction is used to detect the movement of objects at pixel level in image sequences and is an important problem for computer vision. Current methods for solving this task use deep learning models. Compared to traditional methods such as Horn and Schunk, they are significantly faster in inference with comparable performance in accuracy. One of these deep learning methods is Recurrent All-Pairs Field Transforms (RAFT), an architecture that uses a Recurrent Neural Network (RNN) to optimize the optical flow using a Convolutional Gated Recurrent Unit (GRU). Through improvements in modeling, Liquid Time Constant Networks (LTC), a novel RNN model based on neural ODEs has been developed, which avoids known problems of conventional models. LTCs, a subclass of Liquid Neural Networks (LNN), achieve good results in robustness, accuracy and model expressiveness. Since vanilla LTCs have speed limitations due to the use of numerical ODE solvers, a closed form approximation of the LTC equations was created to overcome this disadvantage. This thesis investigates whether this new RNN type can be advantageously used for optical flow prediction using RAFT architecture. The GRU unit is replaced by LNN units and tested in terms of robustness, accuracy, parameter efficiency and speed. GRU shows in the experiments the best accuracy results starting from a minimum inference time and a monotonic increase in accuracy with increasing number of iterations. CfC(pure) has an accuracy advantage at high time requirements, it is memory efficient and has short inference times. CfC(default) shows the best robustness against image noise, but does not achieve the performance of GRU with almost the same number of parameters.