Show HN: MultiMatte, a Promptable Image Background Removal Model

Hey HN, I'm Shreyash from Feyn. We help companies build custom models from their data.

Today we're releasing MultiMatte, a background removal model you can aim with words. Name an object in your image. MultiMatte keeps just that thing, and removes everything else.

Try it out on your images: https://usefeyn.com/multimatte/.

Demo video: https://youtu.be/XZ5BJWAkOjs

MultiMatte is open source. Build with it using our NoBg library https://github.com/feyninc/nobg. Model card here: https://hf.co/feyninc/multimatte

MultiMatte is the second iteration of our background removal models. The first was FeyNoBg, which we released here https://news.ycombinator.com/item?id=49072462.

The big upgrade is promptability. Most models keep all foreground elements when cutting the background. MultiMatte lets you prompt the exact objects you want to keep and remove everything else. For example, If you have an image with a dog and a bowl, you can ask MultiMatte to keep just the dog.

MultiMatte is built on Meta's SAM 3, a concept-promptable detector that already understands phrases. Our biggest change was in masking. Instead of binary masks that classify each pixel as being inside or outside an object, MultiMatte uses alpha mattes that assign an opacity value to each pixel, with respect to an object. This allows us to better represent hair, fur, motion blur, and other kinds of fuzzy boundaries.

Across all our measured benchmarks, MultiMatte shows a step improvement over SAM 3. On DIS5K, S-measure rises from 0.674 to 0.908 (a 34.6% relative gain), and on DUT-OMRON from 0.792 to 0.901 (13.7%).

Our blog covers more of the training details and results https://usefeyn.com/blog/multimatte.

Happy to answer any questions!

▲

peterldowns 2 hours ago | parent | next [-]

This is awesome. Does the demo run fully client-side? This is at least as good as my current solution (Preview.app -> Tools -> Remove Background), I generally pre-crop so it's not too hard to get just the object I care about.

	▲	snyy an hour ago \| parent [-]
		Demo runs on an AWS VM with L4 GPUs. We don't record any logs/data, so feel free to use the demo on whatever images you want. MultiMatte is open source, so you can also run the model locally or in your VM too

▲

aghilmort 2 hours ago | parent | prev [-]

very cool! can we also just oultine trace / approx as hints on input image (or feed it two images, say raw and markup image with whatever visual hints use?

	▲	not_lain 2 hours ago \| parent [-]
		If you mean like blur some parts of the image or draw a contour as preprocessing step then yes. the model does come with bounding box capability but we did not optimize for it in this release. In the next releases we intend to tackle that along with video capabilities and some other features.