Cut anything out of your own photo.
The Segment Anything demo used embeddings we computed ahead of time. This one doesn't: Meta's EfficientSAM — a 2023 FAIR distillation of SAM with a ViT-Tiny encoder — reads your photo and answers your clicks inside this tab, via WebAssembly. nothing is uploaded
Private by construction: the photo is read by your browser and never sent anywhere. The model files download once, when you pick your first photo; after that you can switch off Wi-Fi and it keeps working. Big photos are shrunk to 1024 px on the long side before the model sees them.
Keyboard: focus the picture, then ←↑↓→ move the crosshair (hold Shift for fine steps), Enter places a point, B starts a box and a second B finishes it, Backspace undoes and Esc clears. Tab leaves the picture; the page scrolls again once you do.
Tap or click the thing you want; add up to six points to grow the selection, or drag a box around it · move the mouse to preview before clicking. Or work it from the keyboard — the line under the buttons says how. Then cut it out and save the sticker. (EfficientSAM has no "remove" points — clear and try again if it grabs too much.)
About this demo
parameters, both halves
Counted from the two ONNX graphs this page downloads: 6.16 million in the image encoder, 4.06 million in the prompt decoder. EfficientSAM-Ti replaces SAM's ViT-H image encoder — 632 million parameters, as the EfficientSAM paper counts it — with a ViT-Tiny small enough to run in WebAssembly on a phone.
how it learned
"SAM-leveraged masked image pretraining": the small encoder is trained to reconstruct the features of SAM's big encoder from masked-out image patches, then fine-tuned on SA-1B. Paper: Xiong et al., CVPR 2024.
in your tab
The official ONNX encoder (24.8 MB) and decoder (16.6 MB) from Meta's repo, 8-bit quantized on a Raspberry Pi to 6.6 MB + 4.9 MB, served compressed — plus the ~3 MB WebAssembly runtime. On our test photo its three candidate masks agree with the full-precision model at 0.93, 0.97 and 0.999 IoU, and the embeddings match at 0.997 cosine.
fully open
Code and weights are open source. It sits in the Segment Anything family on the Universe map, alongside SAM 2, SAM 3 and SAM 3D.