Hi @AustinCStone, thanks for sharing your SMURF code! I am trying to port the SMURF Kitti weights to PyTorch. Do you know if your RAFT implementation entails any different architectural design choices from the TorchVision RAFT implementation (except for the use of Instance Normalization in both encoders)? Do you have a script to convert the checkpoints to PyTorch? Thanks for your help :)