MultiMatte

MultiMatte, released by Feyn Labs (Feyn Research), is an open-source background-removal model that can be aimed with a plain-English phrase: tell it "the dog" or "the jeans" and it keeps only that object, producing a smooth alpha matte instead of a hard-edged cutout. It's built on Meta's SAM 3 via low-rank (LoRA) fine-tuning that updates only 19.49M of SAM 3's 860M parameters (2.27% of the weights), yet substantially improves segmentation quality — scoring a 0.901 S-measure on the DIS-VD benchmark versus SAM 3's 0.667, with similar gains across eleven other benchmark splits. Unlike SAM 3's binary inside/outside masks, MultiMatte assigns each pixel a continuous opacity value, which is what lets it handle fine, translucent detail like hair or a blurred screen edge that boolean masks lose. The model weights are published on Hugging Face (feyninc/multimatte) and run through Feyn's own open-source nobg Python library (pip install nobg), with the LoRA adapter already merged into the released weights so no separate adapter library is needed at inference time. It can also be tried directly in the browser at usefeyn.com/multimatte with no installation.