Model rollouts shown for multiple baselines: Dilated ResNet, Constrained Dilated ResNet, U-Net, TF-Net [Wang et al 2018], Constrained TF-Net [Wang et al 2018], and (for Incomp-2D only) Fourier Neural Operator [Li et al 2020]. All models are trained with noise. All models rollouts start such that the first predicted step is the 7th step of the ground truth trajectory. This is to allow comparison with TF-Net, which takes C=6 context states as input.