SparseSAM: Structured Sparsification of Activations in Segment Anything Models

The Segment Anything Model (SAM) achieves strong open-vocabulary segmentation, but its ViT-based image encoders dominate inference latency and memory. Existing activation compression methods, such as token merging, reduce the token length to process, yet introduce non-trivial runtime overhead and encounter catastrophic quality drop under high compression. Other methods applying Sparse Attention focus on attention alone, leaving the MLP fully dense and capping achievable speedup. We propose \emph{SparseSAM}, a \textit{(i) training-free structured sparsification} framework that jointly accelerates attention and MLP layers while preserving token identity. SparseSAM introduces \textit{(ii) Stripe-Sort Attention}, which uses a deterministic Z-order permutation to transform dense attention into static hardware-friendly sparse patterns, eliminating dynamic masking overhead. SparseSAM further introduces a \textit{(iii) Residual-Consistency MLP} that routes only informative tokens through the MLP while propagating remaining tokens through the residual pathway. Across four segmentation benchmarks, SparseSAM loses only 0.004 mIoU at a 0.4 density and 0.021 mIoU at 0.3, a 2.10× reduction in accuracy loss versus token merging advances, while achieving 2x faster inference and 2.8× memory reduction.