MultiMatte Free

What is MultiMatte?

MultiMatte is a promptable image background-removal model built by Feyn Labs. Instead of the usual "remove the background" button, you upload a photo and optionally type a short phrase naming the exact object you want to keep — for example "dog" in a photo of a dog and a bowl, or "white truck" in a cropped street scene. MultiMatte then produces a soft alpha matte for that object and removes everything else, following the standard compositing equation I = αF + (1 − α)B rather than a hard cutout mask.

Under the hood, MultiMatte is a fine-tune of Meta's SAM 3 segmentation model: Feyn Labs used low-rank adaptation to update roughly 19.5M of SAM 3's 860M parameters, turning a general segmentation model into one specialized for high-quality matting driven by a text prompt. The model weights are open, and Feyn has published a field note explaining the training approach alongside a free browser demo where anyone can try it on their own images.

Who it is for

MultiMatte is useful for anyone who regularly needs to isolate a specific subject from a busy photo — e-commerce sellers cutting out one product from a cluttered shot, photographers separating a subject from a background with other people or objects in frame, and designers who need clean alpha mattes (soft edges, hair, fur, semi-transparent regions) rather than the harder edges typical background removers produce. It's also relevant to developers and ML teams who want an open-weights matting model they can self-host or fine-tune further, rather than relying on a closed API.

It is less suited to users who just want a one-click "remove background" workflow with no prompt to write, or who need a polished consumer app with batch processing, brand kits, and export presets — for that, a dedicated product like remove.bg is more turnkey.

Key features

• Promptable subject selection — name the object you want to keep in a few words; leave the prompt empty and MultiMatte will guess the main subject automatically
• Soft alpha matting, not hard masking — outputs a proper alpha channel (I = αF + (1 − α)B), which handles semi-transparent and fine detail edges like hair or fur better than a binary cutout
• Built on SAM 3 — inherits Meta's SAM 3 segmentation backbone, fine-tuned with low-rank adaptation on a small fraction of its parameters
• Open weights — the model itself is openly published, not locked behind a proprietary API
• Free browser demo — upload, paste, or drag-and-drop PNG, JPEG, WebP, or HEIC images (up to 4MB) directly at usefeyn.com/multimatte with no account required
• Crop tool built into the demo — crop the input image before matting to focus the model on the right region

Pricing

Feyn's site does not list a paid plan, subscription, or per-image cost for MultiMatte. The web demo is free to use directly in the browser, and the model weights are published as open weights. There is no indication of API pricing or a hosted commercial tier on the vendor's site at the time of writing, so treat this as a free, open research tool rather than a metered commercial product.

Strengths

• Prompt-driven subject selection is genuinely more flexible than fixed "remove background" tools
• Soft alpha output handles hair, fur, and semi-transparent edges more convincingly than hard-mask competitors
• Open weights mean developers can self-host or build on it instead of depending on a black-box API

Weaknesses

• No polished product wrapper — it's a research demo and open model, not a full SaaS with batch tools, presets, or team features
• No published pricing or API terms, so there's no guarantee of commercial support, uptime SLAs, or continued hosting of the free demo
• 4MB upload limit on the browser demo restricts use with very large or high-resolution source photos

Getting started

Go to usefeyn.com/multimatte, drop, paste, or upload an image (PNG, JPEG, WebP, or HEIC, up to 4MB), optionally type a short phrase naming the object to keep, and click to generate the matte. No signup is required for the demo. Developers who want to self-host or fine-tune the model further should read Feyn's field note, which documents the SAM 3 base model and the low-rank fine-tuning approach used to train it.

Verdict

MultiMatte is a sharp, narrow tool: prompt-based, alpha-quality background removal built on top of SAM 3 with open weights. It won't replace a full product like remove.bg for teams that want batch processing and a polished dashboard, but for anyone who needs precise, prompt-controlled subject isolation — or an open model to build on — it's one of the more interesting releases in AI Photo Editing this year.