Feyn's MultiMatte Model Offers Precise, Promptable Background Removal

Feyn has released MultiMatte, an open-source background removal model that allows users to precisely isolate specific objects within an image using natural language prompts. This approach moves beyond traditional foreground/background segmentation by enabling granular control over what remains.

The core innovation lies in its 'promptability'. Instead of simply identifying all foreground elements, MultiMatte lets users name the exact objects they wish to retain. For example, in an image containing a dog and a bowl, one can instruct the model to keep only the dog, removing everything else.

Technically, MultiMatte is built upon Meta's SAM 3, a concept-promptable detector. Feyn's significant enhancement comes from its masking technique. While SAM 3 typically uses binary masks, MultiMatte employs alpha mattes, which assign an opacity value to each pixel. This method excels at representing fuzzy boundaries such as hair, fur, or motion blur, leading to cleaner and more natural-looking cutouts.

The developers claim a substantial performance improvement over SAM 3. Benchmarks show MultiMatte raising the S-measure on DIS5K from 0.674 to 0.908, a 34.6% relative gain. On DUT-OMRON, the S-measure increased from 0.792 to 0.901, a 13.7% gain.

MultiMatte is available as an open-source library called NoBg, accessible via GitHub. Feyn, the company behind the model, specializes in helping businesses build custom AI models from their proprietary data, positioning MultiMatte as both a product and a demonstration of their capabilities.

Most models keep all foreground elements when cutting the background. MultiMatte lets you prompt the exact objects you want to keep and remove everything else.

For professionals, this model offers a path to automating highly precise image editing tasks. It could significantly reduce the manual effort involved in preparing visual assets for marketing, e-commerce, or design, where isolating specific items with clean edges is critical.

While the technical advancements are notable, its practical utility for Tokyo-based professionals will depend on the robustness of its Japanese language prompting capabilities and its eventual integration into local design or content creation workflows.