Feyn Releases MultiMatte, a Promptable Background-Removal Model You Aim With Words
Feyn's follow-up to FeyNoBg lets you name the object to keep and mattes out everything else. It's open source, built on Meta's SAM 3, and uses alpha mattes so hair and fuzzy edges survive the cut.
2026-09-10 · by Moe Ameen
Published 2026-09-10 · Updated
Short answer
MultiMatte is an open-source background-removal model from Feyn, published September 10, 2026, that you aim with words. Name the object to keep and it mattes out everything else, using alpha mattes so hair and fuzzy edges survive. It is a LoRA fine-tune of Meta's SAM 3 and scores 0.901 S-measure on DIS-VD against SAM 3's 0.667. Weights are on Hugging Face under Apache-2.0.
Key takeaways
- Feyn published MultiMatte, a promptable background-removal model, on September 10, 2026.
- You name the subject to keep; alpha mattes keep hair and fuzzy edges clean.
- It fine-tunes about 2.27% of SAM 3's weights and scores 0.901 S-measure on DIS-VD versus SAM 3's 0.667.
- Weights are at feyninc/multimatte on Hugging Face under Apache-2.0, with a demo at usefeyn.com/multimatte.
- The write-up covers still images only, with no stated video support.
What is Feyn's MultiMatte model?
On September 10, 2026, Feyn published MultiMatte, an open-source model for promptable background removal, in a technical write-up by Hafedh Hichri and Shreyash Nigam. The pitch is simple: it's a background remover you aim with words. Point it at a photo, name the object you want to keep, and MultiMatte isolates that subject and mattes out everything else. It's the second model in Feyn's background-removal line, following FeyNoBg in July 2026; the new capability is selective, prompt-driven object retention rather than keeping every foreground element.
Technically, MultiMatte is built on Meta's SAM 3, the concept-promptable detector released in 2025, and inherits its text alignment and vocabulary. Feyn applied low-rank fine-tuning (a rank-16 LoRA adapter) that modifies about 19.49M of SAM 3's roughly 860M parameters — around 2.27% of the weights — targeting attention and MLP projections, including the CLIP text tower. The key difference from standard segmentation is the output: instead of a binary keep/drop mask, MultiMatte produces an alpha matte that assigns a continuous opacity value to every pixel, so fuzzy boundaries like hair, fur, and motion blur come through as they actually look rather than as a hard-edged cut.
Feyn reports large accuracy gains from that tuning. On the DIS-VD benchmark it posts a 0.901 S-measure versus SAM 3's 0.667, with improvements ranging from roughly +0.039 to +0.244 S-measure across the twelve benchmark splits it reports. The model was trained on 19,953 images over 14,000 steps using focal plus Dice loss, with human-written labels covering 4,949 of those images for prompt supervision.
It ships openly. MultiMatte is distributed through Feyn's NoBg library (Apache-2.0), with model weights on Hugging Face under `feyninc/multimatte` and a browser demo at usefeyn.com/multimatte. The write-up centers entirely on single uploaded photos, with no stated video support, so today it's best treated as a still-image tool. Treat the specific figures here as a snapshot from the launch write-up and check Feyn's pages for current details.
Why does promptable background removal matter for creators?
- Promptability is the real upgrade. A normal remover keeps whatever it thinks is the foreground; MultiMatte lets you specify the exact subject in a cluttered shot — one product on a busy table, one person in a group — which is the case that trips up cheaper cutout tools.
- Alpha mattes matter for anyone who shoots real subjects. Continuous opacity per pixel means hair, fur, and motion blur cut cleanly instead of leaving a halo, which is the difference between a usable product or portrait cutout and one you have to redo.
- It's open and free to self-host. Weights on Hugging Face plus an Apache-2.0 library mean no per-image fee and no cloud upload — but also that it's a model and a library, not a polished consumer app, so getting volume out of it means code.
- It's still-image only for now. The write-up doesn't address video support, so removing a background from a clip means running it frame by frame through your own pipeline, not a one-click export.
- A cutout is still just raw material. MultiMatte hands back an isolated subject on transparency; turning that into captioned, on-brand posts across platforms is a separate job it doesn't touch.
How do you turn a MultiMatte cutout into posts with Kompozy?
If you sell or promote products, the way to act on this today is to treat one promptable cutout as the seed of a launch campaign, not a one-off asset. Photograph the product however you can — a messy desk, a busy shelf — and use MultiMatte's prompt to isolate exactly the item, hair-clean edges and all. That transparent cutout is where Kompozy picks up. Drop it in as the subject of a brand-exact Carousel walking through the product's features, a Photo Post with your hero background, and a Quote Graphic pulling a review line over it — HyperFrames composites the subject onto your styling and layers on-brand captions and offer text, then reframes each piece to 9:16, 1:1, and 16:9 for every destination.
The concrete play for a product drop is turning that single cutout into a week of coordinated posts instead of one image. Kompozy fans the same subject across Photo Posts, Carousels, and Quote Graphics, drafts the supporting Text Posts, Blog Article, and Email Newsletter in your voice through a Persona Brief, and then Autopilot plus a per-post review pipeline schedule and publish the whole batch across the eight social platforms plus blog and email while attention is highest. MultiMatte does the hard part of the matte — picking one subject out of a busy frame; Kompozy does the hard part of the launch — making and shipping the content around it.
What are the quick takeaways on MultiMatte?
- Feyn released MultiMatte — a promptable, open-source background-removal model — on September 10, 2026.
- You name the object to keep; it mattes out everything else, using alpha mattes for clean edges on hair and fuzzy boundaries.
- Built on Meta's SAM 3 via a small LoRA fine-tune; open weights on Hugging Face (feyninc/multimatte), Apache-2.0 NoBg library, demo at usefeyn.com/multimatte.
- Still-image only for now — the write-up doesn't address video support.
- A cutout is raw material; pair it with Kompozy to composite, caption, and publish the subject as finished posts.
What does this look like in Kompozy?
Real Kompozy output: a photo post from the workspace, shown as an example of the format rather than a post about this topic.
Who is this for, and who should skip it?
Best for:
- Product sellers and photographers who need one subject cut out of a cluttered shot
- Developers who can run an open model through the NoBg library
Not ideal for:
- People who want a one-click consumer app; this is a model and a library
- Anyone who needs video background removal, which the write-up does not cover
Frequently asked questions
What is MultiMatte?
MultiMatte is an open-source, promptable background-removal model from Feyn, published September 10, 2026. You name the object you want to keep and it isolates that subject and mattes out the rest. It is built on Meta's SAM 3 with a small low-rank fine-tune and outputs alpha mattes (a per-pixel opacity value) so fuzzy edges like hair survive the cut.
How is MultiMatte different from FeyNoBg?
FeyNoBg, Feyn's earlier model, keeps all foreground elements when it removes the background. MultiMatte adds promptability: you specify which object to keep by naming it, so it can isolate one subject out of a cluttered scene rather than retaining everything in the foreground.
Is MultiMatte free and open source?
Yes. MultiMatte ships through Feyn's NoBg library under the Apache-2.0 license, with model weights on Hugging Face (feyninc/multimatte) and a browser demo at usefeyn.com/multimatte. Your only cost is your own compute and integration time. Check Feyn's pages for the exact terms before commercial use.
Does MultiMatte work on video?
Not directly. It is a still-image model — the write-up doesn't address video support. Removing a background from a clip today means running the model frame by frame through a pipeline you build.
How do we know this?
We checked every date, figure and product claim on this page against the sources listed below on 2026-10-23, and rewrote or removed anything they did not support. Model figures come from Feyn's write-up; the license comes from the Hugging Face model card and the NoBg package on PyPI, and the SAM 3 date comes from its arXiv paper. Kompozy plan and format details come from kompozy.io/pricing.
Sources
- Feyn, MultiMatte (read 2026-10-23)
- Hugging Face, feyninc/multimatte (read 2026-10-23)
- PyPI, nobg (read 2026-10-23)
- arXiv, SAM 3: Segment Anything with Concepts (read 2026-10-23)
- Kompozy, Pricing (read 2026-10-23)