SAM3 vs Astra for text-prompt segmentation: sharper masks vs better language understanding

kate_saenko_ · x · 2026-09-18

A hands-on comparison of SAM3 and Astra for text-prompt segmentation, now rolled out in the Roboflow playground: SAM3 produces much more precise masks but often fails on complicated text prompts, while Astra excels at language understanding and rarely misses the target (mask returned as point-by-point JSON), though it is slower and less accurate than SAM3. Frontier models keep improving at object segmentation, but speed remains Astra's weakness.

Original post →

More from Models

Models channel →