Show HN: MultiMatte, a Promptable Image Background Removal Model

MultiMatte is a new background removal model built on Meta's SAM 3 that uses text prompts to selectively keep objects you name while removing everything else. Using low-rank fine-tuning to modify just 2.27% of SAM 3's parameters, MultiMatte improves segmentation by using alpha mattes instead of binary masks, allowing it to handle fuzzy boundaries like hair and translucent elements. On the DIS-VD benchmark, MultiMatte achieves a 0.901 S-measure compared to SAM 3's 0.667, and is available for testing at usefeyn.com/multimatte.