Particle filters are popularly adopted when state-estimating non-linear stochastic dynamic systems. These filters consist of an ensemble of weighted particles undergoing a sampling procedure divided into two steps: a propagation step driving the ensemble according to a proposal mapping, and a correction step assessing the ensemble likelihood based on observations. Especially in large-scale dynamic systems, the recursive nature of the algorithm causes the ensemble to degenerate after a few iterations. To mitigate this, some filters adopt resampling strategies and informative proposal mappings. However, the resulting ensemble still suffers from the sample impoverishment issue of many particles sharing the same value after resampling. To overcome this, we propose repositioning the predicted particles by a law of motion: a particle flow. Further, we exploit the decay-of-correlations property in large-scale systems through local particle filters which sample in lower-dimensional partitions of the state space. Combining these two approaches removes the need for resampling and alleviates the effects of weight degeneracy and sample impoverishment, allowing for a larger effective number of particles for state estimation (provided that localisation errors and approximation biases remain sufficiently controlled). We illustrate the effectiveness of the resulting local particle flow filter in two systems: one representative of atmospheric dynamics and one modelled as a chemical reaction–diffusion system.