ai-text-to-annotations

ai-text-to-annotations generates polygon or mask annotations from text prompts using SAM3.

v0.2.0 (2026-04-14): New command.

Usage

Provide text prompts describing the objects you want to annotate. The tool generates polygon or mask annotations using SAM3, with image embedding caching for efficient multi-label processing.

labelmetk ai-text-to-annotations [OPTIONS] FILE_OR_DIR

• --texts: Names of objects to annotate. You can specify multiple names separated by comma (e.g., --texts "person,car,building"). Required.
• --model: Backend model. (Default: sam3)
• --shape-type: Output annotation type. One of:
  • polygon (default)
  • mask
• --score-threshold: Threshold to remove annotations by score. The higher value means less annotations in the final result. The range is from 0 to 1. (Default: 0.1)
• --iou-threshold: Threshold to remove overlapping annotations via cross-label NMS. The lower value means less annotations in the final result. The range is from 0 to 1. (Default: 0.5)
• --max-annotations: Max number of annotations. (Default: 100)

Example: Generate polygons for a JPEG

labelmetk ai-text-to-annotations dogs.jpg --texts dog
labelme dogs.export  # open and tweak the annotations

Example: Multiple text labels with mask output

labelmetk ai-text-to-annotations image.jpg --texts "person,car,building" --shape-type mask

Example: Annotate a directory of images

labelmetk ai-text-to-annotations images/ --texts "person,car"
labelme images.export  # open and tweak the annotations