Answers are grounded in this paper and cite their sources.
Resource-constrained vision sensor systems must balance promptable concept detection, segmentation quality, and inference cost. We evaluate structured pruning and grid-preserving token-resolution restructuring of pretrained SAM 3 under prompt-count-aware timing. Historical 500-image detection results describe recovered operating points with different training histories. A separate three-arm, three-seed study controls a short 1000-step recovery budget. On all 5000 COCO validation images, native-mask AP is 0.4714 for the baseline, 0.2168 for Grid36-B6 and 0.0831 for Grid24-B4, so these points do not preserve segmentation quality. On this mask metric, the historical pruned checkpoints rank higher (0.3630 and 0.3109), whereas the short matched-recovery control orders the arms the other way. Five independent paired timing sessions distinguish cached-text box calls from uncached-text mask calls. Historical prompt-path optimizations show why trunk acceleration alone does not predict high-prompt-count latency. Label-recovered pruning is sensitive to prompt phrasing. Further studies compare one INT-FUP pooling adaptation, simulated multi-turn image refinement, and native dynamic FP8 in vision MLPs; each retains explicit quality and runtime limits. Five compiled edge artifacts were measured repeatedly on device, but upstream build effects were not isolated; one artifact received unnormalized input, and incomplete matching-artifact accuracy, including a baseline INT8 collapse, prevents accuracy-preserving edge claims. The evidence supports workload-specific detection/runtime trade-offs rather than general compression superiority. Real video and concepts unseen throughout model development remain unvalidated.
No takes yet. Share an insight, caveat, or question.
Lee et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: