Feyn: MultiMatte brings text-prompted alpha matting to product photography
Feyn released MultiMatte on September 10, 2026, a background removal model that accepts text prompts describing what to keep rather than requiring selection tools or manual masking. The approach is built on Meta’s SAM 3 architecture, extended with alpha matte generation—output values on a 0–255 scale per pixel rather than the hard binary masks that most automated cutout tools produce.
The practical difference shows up at edges. Alpha mattes preserve partial transparency in areas like hair, fabric weave, and semi-translucent materials, which binary masks flatten into hard cuts or jagged outlines. For product photography and e-commerce compositing, where clean cutouts are placed against new backgrounds, the edge quality directly affects how convincing the result looks at close inspection.
On the technical side, Feyn reports S-measure scores of 0.893–0.923 for MultiMatte, compared to 0.649–0.703 for the underlying SAM 3 model. The model was trained on 19,953 images, with 24.8% annotated using human-written text descriptions—the text-prompt capability is a trained behavior rather than a wrapper applied after the fact.
For designers working on product imagery, the relevant change is the removal of manual selection steps for subjects with complex edges. Describing a subject in a prompt—“keep the jeans,” “isolate the bottle”—and receiving a clean alpha-matted cutout reduces the number of passes needed before an asset is ready for compositing. The model handles cases where selection tools typically struggle: fine fabric textures, overlapping elements, and subjects without clean contrast against the background.
Whether this shifts daily workflows depends on the volume of cutout work and the complexity of subjects involved. For studios processing large batches of product images, or for work where edge quality on textiles or hair is a recurring problem, MultiMatte addresses a specific friction point that other automated tools have not closed as cleanly.