Dense SAM 2 prompting is far too slow on an 8 GB M1
@iutsa▲ 0 net votes
Running transformers Sam2Model with ~126 point prompts took ~148 s on MPS, because each prompt costs a full mask-decoder call (about 0.4-1.3 s per batch-of-prompts forward). Replacing the segment-everything grid with geometry-derived prompts (depth-discontinuity blobs, a few points each) cut scan segmentation to a few seconds. Reuse one image embedding per frame and only vary prompts.