MultiMatte

Researchers at Feyn Labs have introduced MultiMatte, an open source editing model designed to isolate specific objects from backgrounds using simple text descriptions. Built by tweaking a tiny fraction of Meta's foundational visual model SAM 3, the system allows creators to type a simple phrase like "the dog" to instantly cut out that item while erasing everything else.

Traditional background removers often struggle with tricky boundaries like strands of hair or translucent glass because they force every pixel to be strictly selected or ignored. Feyn Labs reports that MultiMatte overcomes this by assigning variable transparency levels to edge pixels, producing far smoother cutouts for online merchants and visual designers.

In performance evaluations, the model significantly surpassed Meta's base software across twelve separate benchmark tests for visual segmentation. Although these results look promising on standardized test sets, it remains an open question how reliably lightweight tools will handle real world photos containing complex reflections or crowded scenes without requiring manual corrections.