Feyn Labs has introduced MultiMatte, a background removal model that allows users to specify objects to keep using text prompts. By employing alpha mattes instead of binary masks, the model can describe fuzzy boundaries of translucent or fine elements, such as hair, with continuous opacity values.
Model ReleasesFeyn LabsMultiMatte
Feyn Labs Unveils MultiMatte, a Promptable Model for Precise Image Background Removal
The model is built on Meta's SAM 3. Feyn Labs applied low-rank fine-tuning (LoRA) to modify approximately 2.27% of the 860 million parameters in the original model. This approach allowed MultiMatte to retain SAM 3's existing text-alignment capabilities while significantly improving segmentation performance. On the DIS-VD benchmark, MultiMatte achieved an S-measure of 0.901, compared to SAM 3's 0.667.
The model was trained using 19,953 images covering various scenarios, including salient objects, camouflage, and high-resolution subjects. The training process utilized both focal loss and Dice loss, with prompt supervision provided by 4,949 human-labeled images. MultiMatte is available through the nobg library.
Sources
- Show HN: MultiMatte, a Promptable Image Background Removal Model （Hacker News Frontpage, 2026-09-10）