Show HN: MultiMatte, a Promptable Image Background Removal Model

Feyn released MultiMatte, a promptable image background-removal model built on Meta's SAM 3 that uses low-rank fine-tuning to modify 19.49M of SAM 3's 860M parameters, or 2.27% of the model weights. MultiMatte scores 0.901 S-measure on the DIS-VD benchmark versus SAM 3's 0.667, and improves on SAM 3 across all twelve reported benchmark splits, including 0.979 on DAVIS-S and 0.901 on DUT-OMRON. The model was trained on 19,953 images for 14,000 steps using LoRA rank-16 adapters with focal loss plus Dice loss, and 4,949 images (24.8% of the training set) carried human-written prompt labels.

MultiMatte: Keep What You Want, Cut the Rest We’re introducing MultiMatte, a background removal model you can aim with words. MultiMatte keeps the object you name and removes everything else. Try MultiMatte on your own images at usefeyn.com/multimatte https://usefeyn.com/multimatte . MultiMatte is built on SAM 3 https://huggingface.co/facebook/sam3 Meta, 2025 ref-1 . We used low-rank fine-tuning to modify 19.49M of its 860M parameters. That update touches only 2.27% of the model weights, yet MultiMatte improves substantially on image segmentation. On the DIS-VD benchmark, it scores a 0.901 S-measure against SAM 3’s 0.667. From Segmentation to Matting SAM 3 is a concept-promptable detector. Given a phrase, the model returns binary masks for every matching object. These masks classify each pixel as either inside or outside the object. This binary approach fails on fine or translucent elements with fuzzy boundaries, like hair or a blurred screen. MultiMatte addresses this gap with alpha mattes. Instead of assigning a boolean to each pixel, an alpha matte assigns a continuous opacity value to every pixel. Fuzzy boundaries can then be described as they actually appear. Alpha mattes earn their keep on benchmarks. Across five high-resolution DIS splits, SAM 3 scores between 0.649 and 0.703. MultiMatte scores between 0.893 and 0.923. Training MultiMatte SAM 3 is not one network but several aligned parts: a vision tower, a CLIP text tower, a geometry encoder, a detection encoder and decoder, and a mask decoder. The ability to aim the model with a phrase lives in the alignment between those parts. Our objective was to carry this ability over to matting. We trained the model with PEFT, using LoRA Hu et al., 2021 ref-2 . Every targeted linear layer kept its pretrained weight frozen and learned two small matrices that added a low-rank update alongside it. Training this way let us retain SAM 3’s text alignment and original vocabulary while improving its segmentation ability. The rank-16 adapter followed the configuration principles in Thinking Machines’ LoRA analysis 2025 ref-3 . It targeted the attention and MLP projections in every tower, including the CLIP text tower. We merged the final adapter into the released weights, so inference doesn’t require a separate adapter library. The training run used 19,953 images spanning salient objects, camouflage, high-resolution subjects, hair, and marine scenes. We trained for 14,000 steps with the semantic segmentation objective SAM 3 already uses: focal loss Lin et al., 2017 ref-4 plus Dice loss. For prompt supervision, we used 4,949 images with human-written labels, so 24.8% of the training set named the specific object in each image. This was enough to teach the new matte head to use the text alignment SAM 3 already has. Results We report S-measure, which rewards complete subjects and faithful shapes on a scale from 0 to 1, and mean absolute error over opacity values, where lower is better. Prompt-free S-measure · SAM 3 → MultiMatte | Prompt-free S-measure for SAM 3 and MultiMatte on twelve benchmark splits. | | | | |---|---|---|---| | Split | SAM 3 | MultiMatte | Change | |---|---|---|---| | DIS-VD | 0.667 | 0.901 | +0.233 | | DIS-TE1 | 0.667 | 0.901 | +0.234 | | DIS-TE2 | 0.703 | 0.923 | +0.220 | | DIS-TE3 | 0.685 | 0.921 | +0.235 | | DIS-TE4 | 0.649 | 0.893 | +0.244 | | DAVIS-S no sibling in training mix | 0.913 | 0.979 | +0.066 | | HRSOD-TE | 0.930 | 0.969 | +0.039 | | UHRSD-TE | 0.877 | 0.961 | +0.084 | | DUTS-TE | 0.892 | 0.954 | +0.062 | | DUT-OMRON no sibling in training mix | 0.792 | 0.901 | +0.109 | | COD10K-TE | 0.787 | 0.934 | +0.148 | | CAMO-TE | 0.827 | 0.914 | +0.086 | MultiMatte improves on SAM 3 in every split. We treat changes below 0.002 S-measure or 0.0005 MAE as measurement noise. DAVIS-S and DUT-OMRON are the strongest generalization checks because neither has a sibling in the training mix. MultiMatte reaches its highest absolute score on DAVIS-S at 0.979 and lifts DUT-OMRON from 0.792 to 0.901. Concept names help even without fine-tuning. On DIS-VD, a real concept name adds 0.150 S-measure to SAM 3 with no gradient steps and still adds 0.036 to MultiMatte. The prompt pathway survived the retrain and contributes on its own. CAMO’s named scores use the same annotator’s phrasing style seen during training, so that result measures in-family prompt generalization rather than cross-domain prompting. Run MultiMatte MultiMatte is built with NoBg https://github.com/feyninc/nobg , the same library we used to train FeyNoBg /blog/feynobg . Install it with pip install nobg : python from nobg import AutoModel, AutoProcessor model = AutoModel.from pretrained "feyninc/multimatte" processor = AutoProcessor.from pretrained "feyninc/multimatte" Uses the model's generic default prompt. cutout = model.predict processor, "photo.jpg" cutout.save "output.png" Names the concept to keep. cutout = model.predict processor, "photo.jpg", "the dog" predict returns an RGBA cutout. Pass return type="tensor" for the raw matte. The adapter is already merged into the released weights for ease of use. Try MultiMatte on your own images at usefeyn.com/multimatte https://usefeyn.com/multimatte . MultiMatte and NoBg are built by Feyn. Find us on X https://x.com/feynai , GitHub https://github.com/feyninc , or LinkedIn https://www.linkedin.com/company/107081181/ . Acknowledgements MultiMatte is built on Meta’s SAM 3. FlowDIS Sargsyan and Navasardyan, 2026 ref-5 supplied the human-written DIS5K phrases used for training and evaluation. Thinking Machines’ LoRA analysis informed the adapter configuration. We are grateful to the dataset authors whose released work made the training mix and evaluation possible. arXiv:2511.16719 2025 . arXiv:2106.09685 2021 . arXiv:1708.02002 2017 . CVPR 2026. arXiv:2605.05077 . @note{multimatte2026, title = {MultiMatte: Keep What You Want, Cut the Rest}, author = {Hichri, Hafedh and Nigam, Shreyash and Feyn Research}, year = {2026}, venue = {Feyn Field Notes} }