MultiMatte: Keep What You Want, Cut the Rest
Summary
Overview of MultiMatte, a language-guided background removal model built on SAM 3, employing a low-rank LoRA adapter to enhance text alignment and matting. The article covers the segmentation-to-matting transition, training details, benchmark results showing improvements over SAM 3, and how to run the model with NoBg tooling. It also notes open-source tooling and references foundational papers in the field.