SAM-3 vs SAM-3-LiteText (text-only)
Compare Promptable Concept Segmentation with the same image and text on:
- facebook/sam3 — full SAM-3 text encoder
- yonigozlan/sam3-litetext-s0 — MobileCLIP-based lite text encoder (same ViT-H image encoder, same API)
SAM-3-LiteText is a drop-in replacement: only the checkpoint (and thus text encoder weights) differ. Overlays should be similar; latency on the text path should favor LiteText.
0.05 0.95
0.05 0.95
Examples (COCO val images from the SAM-3 model card)