AgentTested hands-on review
tool: Segment Anything
slug: segment-anything
date: 2026-10-21
score: 7/10

Coverage: One click, one box, and auto-segment. Mug masks compared with rembg.
What we tested: Cut out the FLUX mug from one click and from a box, and auto-segment the hands photo. We compared the mug masks with rembg's cutout as a stand-in reference.
Method: segment-anything (original SAM) with the vit_b checkpoint (375 MB, 1 s download), torch CPU, 8 threads. Point at the mug's centre, box around rembg's mask, SamAutomaticMaskGenerator with 16 points per side. Overlays in sam/contact.png.

Notes:
- CPU-only box: 8 cores, 15 GB RAM (about 4-10 GB free during tests), no GPU.
- 2 images
- rembg mask as reference

Results:
load_s: 1.0
embed_s: 7.0
point: 0.61 s, point [490, 559], scores 0.95, 0.986, 0.918, IoU vs rembg 0.989, 0.981, 0.832, best-scored IoU 0.981
box: 0.14 s, IoU vs rembg 0.982, score 1.005
auto_hands: 27.2 s, 16 points per side, 22 masks
visual: Both mug masks include the handle and exclude the shadow. Auto mode cleanly split the two hands, card stacks and table

Pros:
- Near-identical to rembg's cutout (IoU 0.98) from one click
- After a 7 s image encoding, each new click takes under 1 s
- Apache-2.0, small 375 MB model

Cons:
- Auto-segmenting a whole image takes about 27 s on CPU
- Reference mask was rembg's output, not a hand-drawn one
- Original 2023 SAM; we didn't test SAM 2

Verdict: SAM cut out the mug from a single click almost exactly like rembg did, handle included and shadow excluded. On a CPU it needs about 7 seconds to read an image, then answers each click in under a second. Auto mode split the hands photo into sensible parts in 27 seconds. It's a strong free tool, though we only measured one image.

Score line: Precise one-click masks on CPU. One image measured, against a proxy reference.