An open-source, promptable AI background-removal model — name the object you want to keep and it mattes out everything else.
Last verified · 2026-09-10 · by Moe Ameen
MultiMatte is an open-source model for promptable background removal, released by Feyn — a developer-tools startup — and detailed in a technical write-up published on September 10, 2026 by Hafedh Hichri and Shreyash Nigam. It is the follow-up to Feyn's earlier FeyNoBg. The defining feature is that you aim it with words: instead of keeping whatever it decides is the foreground, MultiMatte lets you name the object you want to keep — the classic example is a photo with a dog and a bowl where you tell it to keep just the dog — and it isolates that subject and mattes out the rest. That makes it genuinely useful on cluttered, real-world shots where "the foreground" is ambiguous.
Under the hood it is built on Meta's SAM 3, the concept-promptable segmentation model from 2025, and inherits SAM 3's text alignment and vocabulary. Feyn's contribution is a small low-rank fine-tune — a rank-16 LoRA adapter modifying roughly 19.49M of SAM 3's ~860M parameters (about 2.27% of the weights) across attention and MLP projections, including the CLIP text tower. Crucially, MultiMatte replaces SAM 3's binary segmentation masks with alpha mattes: every pixel gets a continuous opacity value, so fuzzy boundaries like hair, fur, and motion blur are described as they actually appear instead of chopped to a hard edge. Feyn reports the tuning drives a large accuracy jump — for example a 0.901 S-measure on the DIS-VD benchmark against SAM 3's 0.667.
It is released openly. MultiMatte ships through Feyn's NoBg Python library under the Apache-2.0 license, with model weights on Hugging Face under `feyninc/multimatte` and a browser demo at usefeyn.com/multimatte, so you can try one image or self-host it in your own pipeline.
Two honest notes. First, MultiMatte is a still-image model today — the write-up and demo are both built around a single uploaded photo, with no stated video support, so removing a background from a clip means running it frame by frame yourself. Second, it is a model and a library, not a consumer app: the demo takes a single image, and doing anything at volume means code. Treat the specific numbers here as a launch-write-up snapshot and check Feyn's pages for current details.
MultiMatte's edge over an ordinary remover is precision by prompt: it will pull one named subject out of a messy frame — the founder out of a busy office, one SKU off a crowded shelf — with hair-clean edges. That precision is exactly what makes it a great feeder for the image formats where the subject has to be perfectly isolated before anything else works. Take a MultiMatte cutout into Kompozy as the subject of a Persona Photo, a Photo Post, a Quote Graphic, or a brand-exact Carousel: HyperFrames composites your isolated subject onto your brand background, layers on-style captions and offer or price text, and reframes each piece to 9:16, 1:1, and 16:9 so a single clean cutout becomes native art for every platform.
Where this pays off most is real, imperfect source photos. UGC, phone shots, and event photos rarely have a tidy single-subject foreground — MultiMatte's promptable keep lets you specify the one thing that matters, and Kompozy turns that into a full unit of content: the subject arranged across a Carousel of feature slides, a Quote Graphic pulling a testimonial over it, plus Text Posts and short caption threads written in your voice through a Persona Brief. Then Autopilot and a per-post review pipeline schedule and publish the set across the eight social platforms plus blog and email from one queue. MultiMatte nails the hard cutout; Kompozy turns it into a week of branded, captioned, scheduled posts.
MultiMatte is an open-source, promptable background-removal model from Feyn, detailed in a September 10, 2026 write-up. You name the object you want to keep and it isolates that subject and mattes out everything else. It is built on Meta's SAM 3 with a small low-rank fine-tune and outputs alpha mattes — a continuous opacity value per pixel — so fuzzy edges like hair cut cleanly.
A normal remover keeps whatever it judges to be the foreground. MultiMatte is promptable: you specify which object to keep by naming it, so it can isolate a single subject out of a cluttered scene rather than retaining every foreground element. That control is its main advantage on busy, real-world photos.
Yes. MultiMatte ships through Feyn's NoBg library under the Apache-2.0 license, with 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.
Not directly — the write-up and demo are both built around a single uploaded photo, with no stated video support. Removing a background from a clip today means running the model frame by frame through a pipeline you build.
MultiMatte gives you a clean, prompt-isolated cutout but does not publish it. Bring the transparent PNG into Kompozy to build a Persona Photo, Photo Post, Quote Graphic, or brand-exact Carousel on top of it — HyperFrames composites it onto your brand background with captions — then reframe per platform and schedule across eight social platforms plus blog and email from one queue.