sam2-hybrid · v0.1

Click anywhere on the image. The model is in your tab.

The full sam2.1_hiera_large decoder (224M parameters), running in the browser via onnxruntime-web. Encoder ran separately on a notebook, embedding shipped down once. Every click after that is local inference, no server round-trip.

Source

·
Loading…
left-click — includeright-click — exclude·ONNX · WASM · pre-encoded embeddings

One encode, many clicks

The encoder is heavy and runs once per image. The decoder is tiny (16 MB) and runs once per click. Two halves of the same model on two halves of the network.

Image stays local

You only ship 16 MB of float16 features down to the browser. The original photo never leaves your machine — useful for privacy-sensitive workflows.

Bring your own bundle

Run the Colab notebook to encode any image, then drop the resulting embedding.bin + manifest.json + preview.jpg onto the source picker above.