Show HN: MultiMatte, a Promptable Image Background Removal Model (usefeyn.com)

MultiMatte has been unveiled as a groundbreaking model for image background removal that allows users to specify which object to retain through natural language prompts. Built on the architecture of Meta’s SAM 3, MultiMatte employs low-rank fine-tuning to enhance its capabilities, leveraging only 2.27% of SAM 3’s 860 million parameters to significantly improve image segmentation performance. The model achieved an impressive S-measure score of 0.901 on the DIS-VD benchmark, far exceeding SAM 3’s score of 0.667, showcasing its superior ability to handle complex object boundaries and textures. The innovation of MultiMatte lies in its use of alpha mattes instead of traditional binary masks, which are less effective for intricate details like hair or translucent objects. By assigning a continuous opacity value to each pixel, MultiMatte preserves the subtle variations that characterize fuzzy edges. This capability not only results in visually richer outputs but also indicates a promising advancement in segmentation technology. The model was trained on nearly 20,000 images with a focus on salient objects, employing a hybrid approach that maintains SAM 3’s alignment with text prompts, thus improving usability in real-world applications. Users can experience MultiMatte's capabilities by trying it out at usefeyn.com/multimatte.