NURL objdet — tiny-YOLOv2 on your GPU, pure NURL via WebGPU

The whole detector runs in this tab on your GPU as WGSL compute shaders — the pure-NURL ONNX runtime (packages/onnx + the gpu package's WebGPU backend) compiled to wasm. No server inference, no Python, no OpenCV, no inference engine.

🎥

Start camera or drop an image here.

Needs a WebGPU browser (Chrome/Edge 113+, or Safari 18+).

What you are looking at: a browser tab running a real object detector with zero server compute. The frame is drawn to a 416×416 canvas, the tiny-YOLOv2 network (9 convolutions) runs as WGSL compute shaders on your GPU through NURL's own ONNX runtime, and the YOLO grid is decoded into boxes — all pure NURL compiled to WebAssembly. The 61 MB model is fetched once from the ONNX model zoo and cached.

Slow? Check the adapter. The status line above shows which WebGPU adapter the browser handed over. If it warns about a software adapter (SwiftShader), the shaders are running on the CPU — on Linux, Chrome does this by default. To use the real GPU: open chrome://flags, set Vulkan (#enable-vulkan) and Unsafe WebGPU Support (#enable-unsafe-webgpu) to Enabled, and relaunch the browser. Windows and macOS use the real GPU out of the box. ← playground