SAM3 AI-Box mode, one prompt for a whole crowd

The point of AI-Box has always been "click once instead of drawing the polygon by hand." That's a real speedup when the box contains one object. When the box contains forty cars, the speedup goes away. You either accept a single mask that wraps the whole cluster, or you drag forty separate boxes.

LabelMe v6.1 adds SAM3 AI-Box mode, where one box returns one shape per instance.

LabelMe v6.1 SAM3 AI-Box mode producing multiple shapes from a single bounding box

How it works

Pick the SAM3 model in the AI-Box dropdown. AI-Box ships with multiple SAM backends; SAM3 is the one that understands "find every instance in this region." Older SAM variants are still there for the single-mask case.

Drag the box around the cluster. The model receives the box as a region prompt and returns one mask per instance it detects inside. Each mask becomes its own shape on the canvas.

Each instance gets the same label. Every shape produced by a single SAM3 prompt shares the label dialog you pick once, so a parking-lot box becomes forty car shapes in one step.

When to reach for it

Aerial parking lots and motor pools. The canonical case. Every car is the same class, every box you'd draw by hand is identical work.
Retail shelves. Identical SKUs lined up, each one needing its own bounding shape.
Microscopy. Cells, colonies, particles in a frame. The frame contains a hundred instances; you'd rather not click a hundred times.
Stadium crowds, herds, schools. Anywhere "many instances of the same thing in one region" is the recurring case.

For a single, isolated object, the older AI-Box modes are still the right tool. SAM3 is for the case where one box ought to do the work of many.

Pairs well with

AI-assisted annotation for crowd objects, shipped later in v6.3, suppresses the duplicates that fall out of SAM3 when nested masks overlap or a second pass re-prompts the same region. The two together make SAM3 in a dense scene feel like a clean, additive workflow.

Where to find it

Update to LabelMe v6.1 or later from the download page. SAM3 sits in the AI-Box model dropdown; the older modes are unchanged.

See the v6.1 release notes on GitHub for the full changelog. SAM3 AI-Box landed in #1917, with #1918 gating the unsupported AI-Points combination and #1919 cropping the input for large images.

If a prompt returns nothing or returns the cluster as a single mask instead of separated instances, open an issue with the image and box or ask on Discord.