MultiMatte lets you name the subject to keep in an image instead of relying on automatic background removal to choose it. Feyn introduced the model on October 4 with a browser demo and weights on Hugging Face.
The output is a cutout with transparency. MultiMatte builds on Meta’s SAM 3 but predicts varying opacity around boundaries, which can preserve softer edges such as hair more naturally than a simple inside-or-outside mask. A prompt such as “the dog” tells it which object to retain.
Feyn reports improvements over SAM 3 on its published image-matting evaluations. The published tests also show that naming the subject improves the result compared with leaving the prompt generic.
The web demo accepts an uploaded or pasted image and an optional text prompt. Developers can run the downloadable model through Feyn’s NoBg library and save a transparent PNG. The released weights already incorporate the trained adapter, so the example workflow does not require loading that adapter separately.