The browser version of Microsoft's ONNX Runtime project. It runs a model exported to ONNX directly on the user's device, with no server: on the CPU through WebAssembly and on the GPU through WebGPU. When the page is served cross-origin isolated, WebAssembly can use several threads.
It is used to run AI models on the visitor's device, with no server. The model runs inside a Web Worker, so the page does not freeze meanwhile and images are not sent to a server.
If the browser offers no WebGPU, the model runs on the CPU with WebAssembly. The result is the same; only the time depends on the device.
WebAssembly needs SharedArrayBuffer to use several threads, which requires the page to be served with COOP and COEP headers.