Web YOLO Benchmark

Browser-based YOLOv8 inference benchmark using ONNX Runtime Web : QuickInfer-web

Features

• Run YOLOv8 object detection directly in the browser
• Support for multiple backends: WASM (CPU), WebGPU, WebGL
• Multiple model sizes: YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, YOLOv8x
• Real-time performance metrics: preprocessing, inference, postprocessing
• Mobile browser inference with camera input support

Usage

Local Development

npm install
npm run dev

Open https://localhost:5173 in your browser.

Build for GitHub Pages

npm run build:github

The built files will be in the dist directory.

Models

Models are hosted on ModelScope/quick-infer-models to ensure proper CORS support for browser downloads.

Default models:

• YOLOv8n (~12MB)
• YOLOv8s (~43MB)
• YOLOv8m (~99MB)
• YOLOv8l (~167MB)
• YOLOv8x (~260MB)

Deploy Your Own Models

Edit .env or .env.github:

VITE_MODELSCOPE_REPO=your-username/your-model-repo
VITE_MODELS=[{"name":"your-model.onnx","size":"~50MB"}]

Upload ONNX models to your ModelScope repository.

Browser Compatibility

Backend	Chrome	Edge	Firefox	Safari
WASM	✓	✓	✓	✓
WebGL	✓	✓	✓	✓
WebGPU	✓	✓	Partial	✗

WebGPU provides the best performance but requires browser support.

[todo]

• add repo in the web foot
• yolov11\yolov26 support
• sam3 support