We present SMURF, a method for unsupervised learning of optical flow that\nimproves state of the art on all benchmarks by $36\\%$ to $40\\%$ (over the prior\nbest method UFlow) and even outperforms several supervised approaches such as\nPWC-Net and FlowNet2. Our method integrates architecture improvements from\nsupervised optical flow, i.e. the RAFT model, with new ideas for unsupervised\nlearning that include a sequence-aware self-supervision loss, a technique for\nhandling out-of-frame motion, and an approach for learning effectively from\nmulti-frame video data while still only requiring two frames for inference.\n