SparseSAM: Structured Sparsification of Activations in Segment Anything Models

In three linesSparseSAM introduces training-free structured sparsification for Segment Anything Model's ViT encoders. Using Stripe-Sort Attention (Z-order permutation) and Residual-Consistency MLP, it achieves 2x inference speedup and 2.8x memory reduction with only 0.004 mIoU loss at 0.4 density.

Summary generated by Claude — human-verified