Technology
New promptable background removal model MultiMatte adapts Meta's SAM 3
MultiMatte uses fine-tuning to output alpha mattes for image segmentation, according to Hacker News.
The short version
- Developers introduced MultiMatte, a promptable background removal model built upon Meta's SAM 3 architecture.[Hacker News]
- The model employs low-rank fine-tuning on 19.49 million parameters to generate continuous opacity alpha mattes instead of binary masks.[Hacker News]
- Trained over 14,000 steps on 19,953 images, MultiMatte scored 0.901 on the DIS-VD benchmark compared to SAM 3's 0.667.[Hacker News]
Key facts
- MultiMatte is an image background removal model steered by text prompts and built on Meta's SAM 3 from 2025.[Hacker News]
- The model applies low-rank fine-tuning across 19.49 million of SAM 3's 860 million parameters.[Hacker News]
- MultiMatte generates alpha mattes with continuous opacity values per pixel rather than binary masks.[Hacker News]
- On the DIS-VD benchmark, MultiMatte achieved an S-measure of 0.901 compared to 0.667 for SAM 3.[Hacker News]
- Training ran for 14,000 steps across 19,953 images using focal loss and Dice loss.[Hacker News]
What remains uncertain
- Independent verification of the reported benchmark scores and performance across diverse real-world datasets remains unconfirmed.[Hacker News]
Sources
Outlet counts describe coverage, not independent confirmation. Reports may share a wire service or original source.