Feyn Releases MultiMatte, a Promptable Open-Source Background Removal Model
Feyn released MultiMatte, an open-source image background removal model that can be prompted with words to keep a named object and remove everything else. It is built on Meta's SAM 3 and uses alpha mattes.
MultiMatte lets a user name an object in an image. The model keeps that object and removes the rest. In an example Feyn described, an image containing a dog and a bowl can be prompted so that MultiMatte keeps only the dog. Most models keep all foreground elements when cutting the background, according to the post.
The model is built on Meta's SAM 3, which Feyn describes as a concept-promptable detector that already understands phrases. Feyn said its biggest change is in masking. Instead of binary masks that classify each pixel as inside or outside an object, MultiMatte uses alpha mattes that assign an opacity value to each pixel with respect to an object. Feyn said this better represents hair, fur, motion blur, and other fuzzy boundaries.
Feyn reported that MultiMatte shows a step improvement over SAM 3 across all benchmarks it measured. On DIS5K, the S-measure rises from 0.674 to 0.908, a 34.6% relative gain. On DUT-OMRON, it rises from 0.792 to 0.901, a 13.7% gain.
The model is open source. Feyn said users can try it at https://usefeyn.com/multimatte/, watch a demo at https://youtu.be/XZ5BJWAkOjs, build with it using the NoBg library at https://github.com/feyninc/nobg, and find the model card at https://hf.co/feyninc/multimatte. Feyn's blog at https://usefeyn.com/blog/multimatte covers more training details and results.
Feyn describes itself as a company that helps companies build custom models from their data. The earlier FeyNoBg release was linked at https://news.ycombinator.com/item?id=49072462, and the MultiMatte Hacker News discussion was at https://news.ycombinator.com/item?id=49645803.