In traditional computer vision pipelines, images are processed on fixed uniform grids. In contrast, Gao et al. introduce Deformable Grid, a plug-and-play neural network module that improves alignment with high-frequency information content by deforming the regular grid such that the grid edges and vertices align with image boundaries while the topology remains fixed. DefGrid, which can replace standard pooling methods to reduce feature resolution, can be applied at different levels of image processing for downstream tasks including learnable downsampling, unsupervised image partitioning, and interactive object mask annotation.