Feyn Labs launches MultiMatte to keep the object you name and erase the rest

Why it matters

MultiMatte shows the specialist-model strategy in concrete terms: update 2.27% of a foundation model, retain its language interface, and improve a narrow workflow. The commercial test is whether those gains survive independent evaluation with acceptable latency and cost.

Feyn Labs founder and CEO Shreyash Nigam and software engineer Hafedh Hichri released MultiMatte on September 10th, a promptable image-matting model that keeps the object named by a user and removes everything else. A prompt such as "the dog," "the bowl," or "the jeans" can produce three different cutouts from the same photograph.

For Nigam, the release is another demonstration of the specialist-model thesis behind Feyn Labs. Nigam previously worked on on-device AI, advertising personalization, and virtual-machine memory systems at Google, according to Y Combinator's profile of Feyn Labs. He founded the San Francisco operation with CTO Bhavnick Minhas and took it through YC's Spring 2025 batch. Feyn Labs describes its work as training models around a customer's own data and production feedback, with customers retaining ownership of the resulting weights.

MultiMatte gives that pitch a concrete visual example. Feyn Labs started with Meta's SAM 3, retained its ability to find objects from natural-language prompts, and retrained part of the system to produce finer cutouts. Users can test the web demo or run the released weights through Feyn Labs' open-source tooling.

A matte instead of a binary mask

SAM 3 accepts text and visual prompts, then identifies matching concepts in an image. Its segmentation output generally classifies pixels as either inside or outside an object. That works for distinct shapes, but hard edges are a poor representation of hair, fur, glass, smoke, motion blur, and other details that fade into their surroundings.

MultiMatte produces an alpha matte, assigning each pixel a continuous opacity value. A fully retained pixel can remain opaque, an excluded pixel can become transparent, and uncertain boundary pixels can sit between those endpoints. The output is an RGBA cutout that can be placed directly onto another background.

That distinction matters for image-editing workflows. Conventional background removers usually decide what counts as the main foreground subject. MultiMatte lets the user make that decision with a phrase, including when several plausible subjects occupy the same frame. The feature is useful for product catalogs, creative tools, asset preparation, and editing interfaces where selecting an object manually would otherwise require points, boxes, or a brush.

Hichri's background lines up closely with that problem. He describes himself as a Hugging Face Fellow and has contributed to Transformers, PEFT, Unsloth, Gradio, and computer-vision projects including BiRefNet, BEN2, MatAnyone, and RMBG-1.4. MultiMatte combines that open-source model work with Nigam's preference for adapting foundation models into narrower systems whose behavior can be measured against a defined task.

Feyn Labs changed 2.27% of SAM 3

Feyn Labs says it modified 19.49 million of SAM 3's 860 million parameters, or 2.27% of the weights, using low-rank adaptation. The rank-16 LoRA adapter targeted attention and multilayer-perceptron projections throughout the architecture, including the CLIP text tower responsible for part of the alignment between language and images.

The limited update was central to the design. Feyn Labs wanted to improve image matting without wiping out SAM 3's existing vocabulary or its ability to follow object names. Feyn Labs merged the adapter into the released weights, which means developers do not need a separate adapter package during inference.

Training used 19,953 images across salient-object, camouflage, high-resolution, hair, and marine-image categories. Feyn Labs ran 14,000 training steps using focal loss and Dice loss. Human-written object labels were available for 4,949 images, representing 24.8% of the training set. The remaining images still supported the broader segmentation objective.

The relatively small labeled subset suggests Feyn Labs relied heavily on alignment already present in SAM 3. On the DIS-VD benchmark, Feyn Labs says supplying a real concept name increased SAM 3's S-measure by 0.150 without additional gradient updates. Prompts added another 0.036 to MultiMatte after fine-tuning, indicating that the text pathway continued to contribute after the new matte head was trained.

The benchmark gains are large and company-reported

Feyn Labs reports that MultiMatte scored 0.901 on DIS-VD, compared with 0.667 for SAM 3. Across five high-resolution DIS splits, MultiMatte scored between 0.893 and 0.923, while SAM 3 scored between 0.649 and 0.703.

The reported improvement extended across all 12 benchmark splits in Feyn Labs' table. MultiMatte reached 0.979 on DAVIS-S and 0.901 on DUT-OMRON, two tests Feyn Labs presented as stronger generalization checks because related datasets were not included in its training mixture. It also scored 0.934 on COD10K-TE and 0.914 on CAMO-TE.

Those results have not been independently replicated. Feyn Labs' research note also concentrates on output quality rather than latency, GPU memory, hardware requirements, or per-image inference cost. Those operational measurements will determine whether MultiMatte works as a production component or remains most useful as a research release and local editing tool.

The comparison is also deliberately narrow. Feyn Labs benchmarked MultiMatte against the SAM 3 base model, rather than commercial background-removal services or other dedicated matting models. The tests therefore support Feyn Labs' claim that its fine-tune improves SAM 3 on the selected datasets. They do not establish MultiMatte as the best available background-removal system across every deployment setting.

A second pass at background removal

MultiMatte follows Feyn Labs' July release of FeyNoBg and the NoBg training library. FeyNoBg expanded BiRefNet and focused on producing detailed background-removal mattes. MultiMatte adds text-directed selection, allowing one model to extract different subjects from a crowded scene.

Both releases run through NoBg, Feyn Labs' library for training and using background-removal and image-matting models. The library returns a ready-to-save cutout by default and can expose the underlying matte tensor when developers need lower-level control. MultiMatte can be loaded with the Hugging Face identifier feyninc/multimatte.

The release also fits a broader pattern in Feyn Labs' public work. On June 25th, 2026, Feyn Labs published Pulpie, a family of small encoders for extracting the main content from HTML pages, with applications including pre-training data cleaning. Pulpie and MultiMatte address different data types, but both begin with a tightly defined task, tune a model around measurable output quality, and ship the resulting implementation for developers to inspect.

Nigam's founder letter argues that expertise and production data should shape the model a customer runs, instead of remaining temporary instructions sent to a borrowed system. MultiMatte does not prove the customer side of that business on its own. It does show how Feyn Labs intends to work: preserve useful capabilities from a general model, alter a small portion of its weights, and judge the result against a narrower job where improvements are easier to see.