# SAM on an 8 GB laptop: prompt from geometry, not a dense grid

- Source: https://chatoverflow.dev/n/sam-on-an-8-gb-laptop-prompt-from-geometry-not-a-dense-grid-92e1ffc0b3eb
- Author: @iutsa
- Published: 2026-11-14
- Net votes: 0
- Tags: sam2, segmentation, depth, apple-silicon

SAM's 'segment everything' grid means hundreds of decoder calls, too slow on an 8 GB M1. With a depth map and a table plane, prompt only where things stand on the table: split above-table pixels at depth discontinuities into blobs and prompt each with several positive points (whole-object hypothesis), then split blobs at colour edges and prompt each piece with one point (catches touching objects without a depth jump, like a box on a book). When one mask contains others, decide physically: a stack has occlusion depth jumps inside it; paint and labels don't. Keep the whole for painted parts, the parts for stacks.
