SAM 2.1 tiny — ONNX export for Video Redactor

These files are an ONNX export of Meta's SAM 2.1 tiny (sam2.1_hiera_tiny.pt) used by Video Redactor for its Accurate AI tracking mode. Video Redactor downloads them on first use and verifies every file's SHA-256 before loading it; it runs them locally with ONNX Runtime — no video ever leaves the computer.

They are published here only so the app can download them. They are tied to the app's JavaScript glue code (single-object, box-prompted video propagation) and are not a general-purpose SAM 2 runtime.

File	Contents	SHA-256
`enc.onnx`	image encoder (+ mask decoder's high-res feature convs)	`dc0b5a163d35ac94f2f6f9495acd08b07d01727fa9f51d9339aabae85b278aa4`
`dec.onnx`	prompt encoder + mask decoder	`5bae8dcc82db7c640e80791f207c700b0c99ed9455fd3aa44bbfccc5102e8d77`
`menc.onnx`	memory encoder	`db9fefce908f5402f0838f9317545b4029ac496c45d6c54db2a1c552bd00fe43`
`mattn.onnx`	memory attention (fixed 7-frame bank + 16 object pointers)	`c18925fa3e961591dd5ea7fa3dcbcc4f210c779fa4e2d39c57f09f7b84a412cb`
`consts.npz`	constant tensors (position encodings, no-memory embedding, no-object pointer)	`4c7f2efb06b3f1bf79d349762354091ce4cfe78672f07784d8b3d47380db598e`

Total ≈ 168 MB, float32, ONNX opset 17, image size 512.

Licence and modifications

SAM 2 is © Meta Platforms, Inc. and licensed under the Apache License 2.0 (facebookresearch/sam2). This is a modified form of the model: the predictor was split into four ONNX graphs plus a constants file, the memory attention takes a fixed-size memory bank, and the rotary position encoding was rewritten with real arithmetic so it could be exported. The trained weights are unchanged. See NOTICE for the full modification notice.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Extreme-Media/video-redactor-sam2.1-tiny-onnx

Base model

facebook/sam2.1-hiera-tiny

Quantized

(17)

this model