MultiMatte Turns Background Removal Into a Prompt

Feyn released MultiMatte on 10 September 2026, a background removal model you steer by naming the thing you want to keep. Type "keep the jeans" and it returns the jeans on transparency. It is a low-rank fine-tune of Meta's SAM 3, touching 19.49M of that model's 860M parameters, and on the DIS-VD benchmark it scores 0.901 S-measure against SAM 3's 0.667.

The interesting part is not the prompt. It is what comes back. MultiMatte returns an alpha matte, a continuous opacity value for every pixel, rather than the binary in-or-out mask that most segmentation models hand you. That is the difference between a cutout that survives a hair edge and one that does not.

The caveat is in the benchmark table, and it is worth stating before the numbers rather than after: every comparison Feyn publishes is against SAM 3, which is MultiMatte's own base model. That proves the fine-tune worked. It does not establish that MultiMatte beats the free stack most creators already run.

What Feyn actually shipped

MultiMatte is a rank-16 LoRA adapter on top of facebook/sam3, targeting the attention and MLP projections across all towers. It modifies 19.49M of 860M parameters, or 2.27% of the weights. Feyn trained it for 14,000 steps on 19,953 images covering salient objects, camouflage, high-resolution subjects, hair, and marine scenes. Of those, 4,949 images (24.8% of the set) carried human-written text labels, which is what teaches the model to act on a named concept.

The weights are on Hugging Face as feyninc/multimatte, tagged 0.8B parameters, with a standardized 1008 by 1008 input resolution. Distribution is through NoBg, Feyn's background removal library, installable with pip install nobg. This is a genuinely small release by current standards: the model card showed 199 downloads in the trailing month at the time of writing, and the Show HN thread drew 55 points and 8 comments. In that thread a Feyn team member confirmed the demo runs on AWS L4 GPUs, that no logs or image data are recorded, and that video support is planned but not shipped.

Alpha mattes are the point, not the prompt

A binary mask makes one decision per pixel: foreground or background. That works when the boundary is a hard edge, like a phone against a seamless backdrop. It fails on everything soft. Hair, fur, motion blur, smoke, glass, a translucent fabric edge: at those boundaries a single sensor pixel genuinely captured a blend of the subject and whatever was behind it. Forcing that pixel to one side produces either a jagged fringe or a visible halo.

An alpha matte assigns each pixel a continuous opacity instead, so a strand of hair can be 40% subject and 60% background, and it composites correctly onto a new backdrop. This is the same reason compositors have used mattes rather than masks for decades, and it is the practical reason MultiMatte's gains cluster where they do.

One honest limitation applies to every matting model, not just this one: opacity does not unmix color. If a pixel captured a blend of your subject and a bright green wall, marking it semi-transparent keeps the green in the blend, and you can still get a color fringe against the new background. Matting improves the edge shape. It does not undo spill. The same trade-off shows up when you remove a video background with AI, where it compounds across frames.

The benchmark table, and the baseline it leaves out

Feyn reports S-measure across 12 benchmark splits. These are Feyn's own evaluation runs, and the comparison column is SAM 3 throughout.

Read the spread rather than the average. The largest gains land on the DIS splits (+0.220 to +0.244) and on COD10K-TE (+0.148), which is a camouflaged-object set. The smallest land on HRSOD-TE (+0.039), DUTS-TE (+0.062), and DAVIS-S (+0.066), where SAM 3 already scored between 0.892 and 0.930. The pattern is consistent: SAM 3 was competent at large, obvious, well-lit subjects, and the fine-tune bought its improvement on intricate boundaries and low-contrast subjects. That is precisely the product-photography and compositing case, so the gains land where a working creator would feel them.

There is also a revealing ablation. Feyn reports that supplying a real concept name adds 0.150 S-measure to stock SAM 3 with no gradient steps at all, and adds only 0.036 to MultiMatte. Most of the benefit of naming the object was already available in SAM 3; the fine-tune absorbed it into the weights. If you are already prompting SAM 3 by hand, your headroom here is smaller than the top-line 0.233 suggests.

What the table cannot tell you is the comparison that actually decides your tooling. Feyn does not benchmark against BiRefNet or against the rembg distribution that wraps it, and those are the incumbents. Until someone publishes that head-to-head, treat MultiMatte as demonstrably better than its own base model and unproven against the alternative on your disk.

How it compares to what you already run

Since the published numbers do not settle the quality question, the honest comparison is on the attributes you can verify. The free incumbent is rembg, which exposes multiple BiRefNet variants plus other backends through a CLI, a Python API, an HTTP server, and Docker.

The column that matters is the first one. Every other option decides for you which object is the subject, and when a frame holds two plausible subjects you are left refining a selection by hand. Naming the subject is the actual workflow change here, and it is why this is worth attention even before the quality question is settled. It is a different kind of control than the selection tooling in Photoshop's AI assisted editor, which is faster at refinement but still expects you to indicate the subject.

The license line the model card does not draw

The Hugging Face card tags MultiMatte apache-2.0, and the NoBg library is genuinely Apache-2.0. Read quickly, that says unrestricted commercial use. The stack is more complicated, because MultiMatte is 19.49M parameters of adapter sitting on 860M parameters of facebook/sam3, and SAM 3 does not ship under Apache. It ships under Meta's custom SAM License, last updated 19 November 2025.

Three terms in that license are worth knowing before you build on this. First, ownership of your own work is explicit: section 5a states that as between you and Meta, you own the derivative works and modifications you make. Second, the license follows the weights. Distribution of SAM materials "and any derivative works thereof" remains subject to the agreement, and anyone redistributing must include a copy of it. Third, there are acceptable-use carve-outs covering ITAR-controlled activity, military or warfare purposes, nuclear applications, espionage, weapons development, and activity subject to trade controls or sanctions, plus a requirement to acknowledge use of the materials in publications.

For most readers this changes nothing practical. Cutting out product shots for a client, running the model on your own machine, shipping the resulting PNGs: nothing in either license obstructs that. The obligations bite at redistribution, which is the case where you embed the model inside a tool you hand to someone else. At that point your terms come from Meta's license, not from the Apache tag on the adapter. If that is your situation, read the license text against your own facts rather than relying on a summary, including this one.

Running it

The fastest evaluation path is the hosted demo, which takes about a minute and costs nothing.

1. Open the MultiMatte demo and upload a frame you already know is hard: hair against a busy background, a chair with thin legs, or a product with a translucent edge. Feyn states the demo records no logs or image data.
2. Type the subject as a short noun phrase, for example "the dog" or "the jeans", rather than a full sentence describing the scene.
3. Compare the edge against whatever produced your last cutout. Zoom to 200% on the worst boundary in the frame. That is the only comparison that decides anything.
4. If it wins, install the library locally with pip install nobg. It needs Python 3.10 or newer, PyTorch 2.0 or newer, and TorchVision 0.15 or newer.

The model card documents the minimal call:

from nobg import AutoModel, AutoProcessor

model = AutoModel.from_pretrained("feyninc/multimatte")
processor = AutoProcessor.from_pretrained("feyninc/multimatte")
model.predict(processor, "photo.jpg", "the dog").save("dog.png")

The NoBg repository documents the options that matter for production: batch processing, returning the raw alpha matte instead of a composited cutout, and a foreground refinement pass. It also carries a second Feyn model alongside the SAM 3 backend, so the library is worth reading before you wire it into a pipeline. If you are already maintaining local image models, the operational shape is familiar from running a local LoRA trainer: small adapter, large base, and the base is where your constraints live.

Frequently asked questions

Is MultiMatte free to use commercially?

The adapter and the NoBg library are Apache-2.0, and the SAM 3 base ships under Meta's custom SAM License rather than an open-source license. That license carries acceptable-use restrictions and flows through to derivative works when you distribute them. Using it to produce images for clients is a different question from shipping the model inside a product, and only the second case requires you to read the license carefully against your facts.

Does MultiMatte beat rembg and BiRefNet?

Nobody has published that comparison, including Feyn. The 12 benchmark splits Feyn reports all compare MultiMatte to SAM 3, its own base model. Test it against your current tool on your own hardest frames before switching.

What is the difference between an alpha matte and a binary mask?

A binary mask marks each pixel as fully foreground or fully background. An alpha matte assigns a continuous opacity, so a pixel that genuinely captured a blend of subject and background can be partially transparent. The difference is visible on hair, fur, motion blur, smoke, and glass.

Do I need a GPU to run it?

The model is 0.8B parameters and Feyn runs the public demo on AWS L4 GPUs. It will run on CPU through PyTorch, but slowly, in the same way other background removal models degrade to several seconds per image without acceleration. For batch work, a GPU is the practical requirement.

Does MultiMatte work on video?

Not yet. A Feyn team member said in the Show HN thread that video capability is planned for a future release. For now it is a single-image model, so applying it across frames means handling temporal consistency yourself.

What resolution does it work at?

Inputs are standardized to 1008 by 1008, inherited from the SAM 3 architecture. For high-resolution production images that means the matte is computed at that size, so very fine detail in a large frame may need upscaling of the matte or a tiled approach.