Segment Anything in High Quality
lastmod 2026-11-10created 2023-06-272 edits
Segment Anything in High Quality (HQ-SAM) is a follow-up model that improves the boundary quality of the masks produced by Meta's Segment Anything Model (SAM). It keeps SAM's pretrained weights frozen and trains only an added high-quality output token and feature fusion module, so it segments fine structures and thin boundaries more accurately with little data and cost. It can be used with the same prompt types as SAM (points, boxes).
Segment Anything in High Quality
Overview
SAM_vs_HQ-SAM_Overview.png
SAM vs HQ-SAM
SAM_vs_HQ-SAM_Demo_4.gif