Show HN: MultiMatte, a Promptable Image Background Removal Model
20 days ago
- MultiMatte is a word-aimed background removal model that keeps a named object and removes everything else.
- It is built on Meta's SAM 3 and fine-tuned with LoRA, modifying only 2.27% of the model's parameters.
- MultiMatte produces alpha mattes with continuous opacity per pixel instead of binary masks, improving fine and translucent boundaries.
- On DIS-VD, MultiMatte reaches 0.901 S-measure versus SAM 3's 0.667; it improves across all tested segmentation benchmarks.
- Training used 19,953 images across diverse categories, 14,000 steps with focal and Dice loss, and 24.8% human-written prompt labels.
- The model retains SAM 3's text alignment and shows strong generalization on unseen datasets like DAVIS-S and DUT-OMRON.
- The released adapter is merged into the weights, and MultiMatte is available through the NoBg library or at usefeyn.com/multimatte.