Build 1 publisher2 min readPublished September 10, 2026 at 2:46 PM PDT

Naming the object was worth 0.150 to SAM 3 before Feyn Labs trained anything

Feyn Labs' MultiMatte updates 2.27 percent of SAM 3's weights with a rank-16 LoRA, and every pixel in the output now carries a continuous opacity value. The DIS-VD figures supporting it are Feyn Labs' own.

The Engineer · Build desk

What happened

Feyn Labs' Shreyash Nigam and Hafedh Hichri released MultiMatte on September 10th, a promptable matting model that keeps whichever object a user names in a phrase and removes everything else.
MultiMatte assigns every pixel a continuous opacity value and returns an RGBA cutout that can be placed straight onto another background.
Feyn Labs says it changed 19.49 million of SAM 3's 860 million parameters, 2.27 percent of the weights, using low-rank adaptation.
Human-written object labels covered 4,949 of the training images, 24.8 percent of the set, with the rest supporting only the broader segmentation objective.
The adapter is merged into the released weights, so developers load a single checkpoint and need no separate adapter package at inference.

Compiled by The EngineerSomething wrong?How this is made

Why it matters

capability A phrase now picks the subject, so a catalog or editing pipeline can drop the points, boxes and brush strokes an operator used to supply per image.
decision Anyone evaluating a specialist model of this shape has to score the frozen base model with the same prompt first, or the adapter gets credit for alignment it inherited.
constraint Teams that want both base SAM 3 behaviour and MultiMatte behaviour in one process have to host two checkpoints, because the adapter cannot be unloaded.
cost The adopter funds the measurement that decides this: runtimewire puts the commercial test at whether the gains survive independent evaluation with acceptable latency and cost.

The release reports two DIS-VD deltas, and they measure different things. Feyn Labs says supplying a real concept name raised unmodified SAM 3's S-measure by 0.150, with no additional gradient updates ^[16]. After the fine-tune, prompting added a further 0.036 to MultiMatte ^[17]. The first is about four times the second ^[2].

So the base model already mapped "the dog" onto the dog's pixels, and the adapter mostly bought edge quality ^[2]. The training data fits that reading. Human-written labels covered 4,949 images out of 19,953, leaving 15,004 with none ^[15]^[13]^[1]. Feyn Labs ran 14,000 steps on focal loss and Dice loss ^[14].

The adapter is rank-16 and sits on attention and multilayer-perceptron projections throughout the architecture, including the CLIP text tower that carries part of the alignment between language and images ^[10]. That last placement is the one worth arguing about at review. Feyn Labs says the small update was deliberate, because it wanted better mattes without wiping out SAM 3's vocabulary or its ability to follow object names ^[11]. No measurement of retained vocabulary appears in the release. A team with a long product taxonomy would want that test before swapping checkpoints.

SAM 3's segmentation generally sorts each pixel inside or outside an object ^[4]. Hair, fur, glass, smoke and motion blur do not survive that representation, because they fade into what is behind them ^[5]. Hichri, who lists himself as a Hugging Face Fellow, has contributed to BiRefNet, BEN2, MatAnyone and RMBG-1.4 ^[21], so the new matte head comes from someone who has shipped this class of model before.

Feyn Labs went through Y Combinator's Spring 2025 batch and describes its business as training models on a customer's own data and production feedback, with the customer keeping the resulting weights ^[19]^[20]. The weights here ship with the adapter merged in ^[12]. Convenient, right up until you want base SAM 3 back in the same process.

The release does not publish an absolute S-measure or a head-to-head matting score against base SAM 3 ^[3]. For the two deltas to transfer to your pipeline, your images have to look like the five categories in the training mix: salient object, camouflage, high resolution, hair, and marine ^[13]. Your prompts also have to name concepts SAM 3 already knows, because that is where the larger delta came from, and both deltas are Feyn Labs' own ^[16].

What to watch

An independent DIS-VD evaluation reporting absolute S-measure values for MultiMatte alongside other matting models.
Published latency and throughput for the merged checkpoint on named hardware.
Any third-party test of prompt recall on concepts outside the five training categories.

Claim ledger

Ranked by verification strength, evidence, and original report placement.

[1]
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.
ReportedSupportedView cited source
[2]
A prompt such as "the dog," "the bowl," or "the jeans" can produce three different cutouts from the same photograph.
ReportedSupportedView cited source
[3]
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.
ReportedSupportedView cited source
[4]
SAM 3 accepts text and visual prompts and identifies matching concepts in an image; its segmentation output generally classifies pixels as either inside or outside an object.
ReportedSupportedView cited source
[5]
Hard edges are a poor representation of hair, fur, glass, smoke, motion blur, and other details that fade into their surroundings.
ReportedSupportedView cited source
[6]
MultiMatte produces an alpha matte, assigning each pixel a continuous opacity value, and the output is an RGBA cutout that can be placed directly onto another background.
ReportedSupportedView cited source
[7]
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.
ReportedSupportedView cited source
[8]
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.
ReportedSupportedView cited source
[9]
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.
ReportedSupportedView cited source
[10]
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.
ReportedSupportedView cited source
[11]
Feyn Labs wanted to improve image matting without wiping out SAM 3's existing vocabulary or its ability to follow object names; the limited update was central to the design.
ReportedSupportedView cited source
[12]
Feyn Labs merged the adapter into the released weights, which means developers do not need a separate adapter package during inference.
ReportedSupportedView cited source
[13]
Training used 19,953 images across salient-object, camouflage, high-resolution, hair, and marine-image categories.
ReportedSupportedView cited source
[14]
Feyn Labs ran 14,000 training steps using focal loss and Dice loss.
ReportedSupportedView cited source
[15]
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.
ReportedSupportedView cited source
[16]
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.
ReportedSupportedView cited source
[17]
Prompts added another 0.036 to MultiMatte after fine-tuning, indicating that the text pathway continued to contribute after the new matte head was added.
ReportedSupportedView cited source
[18]
According to runtimewire, the commercial test is whether MultiMatte's gains survive independent evaluation with acceptable latency and cost.
ReportedSupportedView cited source
[19]
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.
ReportedSupportedView cited source
[20]
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.
ReportedSupportedView cited source
[21]
Hichri 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.
ReportedSupportedView cited source
[22]
Users can test the web demo or run the released weights through Feyn Labs' open-source tooling.
ReportedSupportedView cited source
[d1]
15,004 of the 19,953 training images, about 75.2% of the set, had no human-written object label.
DerivedSupported1 source— sign in to open
[d2]
The prompt-only S-measure gain on unmodified SAM 3 (0.150) is about four times the prompt contribution measured on MultiMatte after fine-tuning (0.036).
DerivedSupported1 source— sign in to open
[d3]
The quality figures in the release are prompt-sensitivity deltas of 0.150 and 0.036 on DIS-VD; no absolute S-measure value, no head-to-head MultiMatte-versus-SAM-3 matting score, and no latency, throughput or GPU-hour figure is reported.
DerivedContested2 sources— sign in to open

Sources

Publishers with included, body-backed reporting in this cluster.

runtimewire.com
1 article · September 10, 2026
Feyn Labs launches MultiMatte to keep the object you name and erase the rest

Ask Clarity

Have a question this piece left open? The answer comes from its reporting and claim ledger — nothing else.

Try

Clarity analysis

How the coverage leans and what the record actually supports. Read the methodology

Perspective Coverage

How much of the coverage speaks to builders, operators and investors. It measures emphasis, not whether the story is right.

The reusable parts here are the recipe and the artifacts. A rank-16 LoRA on attention and MLP projections, the CLIP text tower included, changed 19.49 million of SAM 3's 860 million weights, and Feyn Labs merged the adapter back in, so inference loads one set of weights. Training ran 14,000 steps over 19,953 images with focal and Dice loss, and only 4,949 of those images carried a human-written object label. In runtimewire's write-up, handing unmodified SAM 3 a real concept name on DIS-VD lifted its S-measure by 0.150 before any gradient update, while after fine-tuning prompts were worth 0.036 to MultiMatte. Most of the language grounding was already in the base model. Weights and a web demo are public, so both numbers can be checked instead of taken on trust.

Reality

Evidence, adoption, hype gap, incentives and confidence, each read independently from the current record on the scale shown beneath it. How these are measured.

Evidence38

0100

All from the vendor

Every number traces to Feyn Labs: the parameter counts, the dataset size, the step count, and all twelve benchmark splits. They are specific and they hold together arithmetically, since 4,949 labeled images out of 19,953 does leave 15,004 without a human-written label. No one outside the company has run the weights on DIS-VD, and the only baseline offered is the model Feyn Labs started from.

Adoption12

0100

Weights up, users unknown

The artifacts exist: merged weights, open-source tooling, a web demo, all dated September 10th. Beyond those, no customer, integration, download figure or deployment appears anywhere in this reporting, and the earlier FeyNoBg release it follows arrives without usage numbers either.

Hype gap

Topics and entities

Follow any of these and your For You feed starts watching them — no settings page required.

Topics

Entities

invest
OpenAI rates GPT-6 Astra capable of hacking hardened systems without human guidance
1 publisher
build
Unsloth's 10% quant claim is really about which machines can run a 27B model
1 publisher
invest
Washington pitches Carolina Principles to G20, urging no new AI rules or bodies
1 publisher
product
OpenAI stops a "significant number" of Astra training runs until cyber gates are met
7 publishers

What happened

Why it matters

What to watch

Claim ledger

Sources

Let Clarity write the post for you.

Topics and entities

Topics

Entities

Related stories

Claim ledger

Let Clarity write the post for you.

Topics and entities

Topics

Entities

Perspective Coverage

Reality

Related stories

Sources