This paper introduces Task-Oriented Flow (TOFlow), a novel motion representation learned in a self-supervised, task-specific manner to enhance video processing tasks. Traditional video enhancement algorithms often rely on optical flow for frame registration, but precise flow estimation is challenging and sub-optimal for specific tasks. TOFlow is designed as an end-to-end trainable network with a motion estimation component and a video processing component, jointly trained to learn task-oriented flow. The authors build the Vimeo-90K dataset, a large-scale, high-quality video dataset for low-level video processing, to evaluate TOFlow. TOFlow outperforms traditional optical flow on standard benchmarks and the Vimeo-90K dataset in three video processing tasks: frame interpolation, video denoising/deblocking, and video super-resolution. The paper also discusses the advantages of joint training and the effectiveness of task-oriented flow in handling occlusions and improving video quality.This paper introduces Task-Oriented Flow (TOFlow), a novel motion representation learned in a self-supervised, task-specific manner to enhance video processing tasks. Traditional video enhancement algorithms often rely on optical flow for frame registration, but precise flow estimation is challenging and sub-optimal for specific tasks. TOFlow is designed as an end-to-end trainable network with a motion estimation component and a video processing component, jointly trained to learn task-oriented flow. The authors build the Vimeo-90K dataset, a large-scale, high-quality video dataset for low-level video processing, to evaluate TOFlow. TOFlow outperforms traditional optical flow on standard benchmarks and the Vimeo-90K dataset in three video processing tasks: frame interpolation, video denoising/deblocking, and video super-resolution. The paper also discusses the advantages of joint training and the effectiveness of task-oriented flow in handling occlusions and improving video quality.