Introduction

Background removal used to mean clicking, brushing, and praying the hair strands survive. Feyn just flipped that ritual into a sentence. MultiMatte is a SAM 3–based model you aim with words: say “the dog,” “the jeans,” or “the bowl,” and it keeps that subject while cutting everything else — as a real alpha matte, not a crunchy binary mask.

For prompt engineers and cutout artists, it feels like handing your eraser a vocabulary.

What MultiMatte Actually Does

Meta’s SAM 3 already understands concept phrases and returns binary masks. Great for “find every mug.” Less great for soft edges: hair, motion blur, frosted glass. Binary labels force every pixel into inside-or-outside, so fuzzy boundaries get chewed.

MultiMatte keeps SAM 3’s text steering and upgrades the output to continuous opacity. Fuzzy edges become soft alphas instead of jagged silhouettes. You can still run it without a prompt (generic salient cutout) or with a named concept when the frame is busy.

Feyn fine-tuned with LoRA / PEFT: about 19.49M of SAM 3’s 860M parameters (~2.27%), rank-16 adapters across attention and MLP projections — including the CLIP text tower — then merged the adapter into the released weights so inference stays simple.

Training used roughly 19,953 images spanning salient objects, camouflage, high-res subjects, hair, and marine scenes for 14,000 steps (focal + Dice). About a quarter of that mix carried human-written concept labels so the matte head learned to use the phrase pathway SAM 3 already had.

The Numbers (S-measure, higher is better)

On dichotomous high-res splits, MultiMatte jumps hard versus SAM 3:

  • DIS-VD: 0.667 → 0.901 (+0.233)
  • DIS-TE1–TE4: gains around +0.22 to +0.24
  • Generalization checks DAVIS-S and DUT-OMRON (no sibling in the train mix) still climb, peaking at 0.979 on DAVIS-S

Prompt names help even before fine-tuning on SAM 3, and they still add a little lift after MultiMatte’s retrain — the text pathway survived.

Why Creators Should Care

This is the missing middle between “segment anything” and “pro matting”:

  • Promptable keep lists — isolate the jacket, not the whole person; the product, not the prop table
  • Hair-friendly alphas — continuous opacity for soft edges instead of cookie-cutter masks
  • Tiny update, big lift — LoRA-scale change on a familiar backbone, not a from-scratch mystery stack
  • Shippable today — web demo plus Hugging Face weights through Feyn’s NoBg library

Pro Tip

When a scene has multiple subjects, name the smallest useful noun (“the left sneaker,” “the red mug”) rather than a vague “subject.” Iterate prompts the way you iterate seeds — MultiMatte’s whole point is that the keep-set is editable English.

Try It

Demo: usefeyn.com/multimatte

pip install nobg

from nobg import AutoModel, AutoProcessor

model = AutoModel.from_pretrained("feyninc/multimatte")
processor = AutoProcessor.from_pretrained("feyninc/multimatte")

# generic salient cutout
cutout = model.predict(processor, "photo.jpg")
cutout.save("output.png")

# name what to keep
cutout = model.predict(processor, "photo.jpg", "the dog")

Weights: feyninc/multimatte on Hugging Face. Write-up: Feyn Field Notes. Built on Meta’s SAM 3; FlowDIS phrases helped train and eval.

Conclusion

Segmentation found the object. MultiMatte lets you point at it in plain language and walk away with a matte that respects the soft edges. Small LoRA, sunny leap for anyone who lives in cutouts, comps, and product stills.

—Aurelia ♡