BackgroundBlur is deliberately built around local inference. The browser downloads the selected model and machine-learning runtime, estimates a foreground mask on your device, and composites the finished image with Canvas APIs. The source photo does not need to be sent to a BackgroundBlur image-processing endpoint for the core workflow.
Xenova/modnet is the fast human-focused engine. It is a Transformers.js-compatible MODNet conversion and is used for portraits, profile pictures, and other images where the main foreground is a person.
Best for: people, hair, shoulders, portrait-style photos.
Ko033/isnet-general-use-onnx is the balanced general-purpose engine. BackgroundBlur applies the model's required per-image min/max mask normalization before compositing, which is critical for a clean full-range alpha matte.
Best for: most products, pets, vehicles, food, furniture, and mixed everyday images.
onnx-community/BiRefNet_lite-ONNX is the opt-in edge-quality engine. It runs only when WebGPU is available because the browser model is substantially larger and more demanding than the fast paths.
Best for: fine hair, complex silhouettes, spokes, thin structures, and images where edge quality matters more than first-run speed.
Portrait and Universal AI attempt GPU acceleration when practical and retain WebAssembly-compatible paths. High Quality AI is explicitly WebGPU-only; if it cannot initialize, the editor falls back to Universal AI instead of leaving the user stuck. Engine requests are versioned so a slower old request cannot overwrite a newer selection.
Your browser still makes network requests to load site assets, analytics scripts, runtime files, and AI model files. Those requests can expose ordinary request metadata to the relevant infrastructure. The selected source image is not attached to the model-download request and is processed locally by the core editor.
Transparent glass, motion blur, reflections, thin wires, mesh, smoke, and foreground colors that nearly match the original background can produce imperfect masks. The editor therefore includes edge softness plus a manual Restore/Erase brush so an automatic result is not the final word.
The Transformers.js MODNet conversion and Universal ISNet ONNX package are published with Apache-2.0 licensing metadata. BiRefNet Lite's ONNX package is published under MIT. Transformers.js itself is Apache-2.0. BackgroundBlur keeps third-party notices in the source distribution and does not use commercially restricted BRIA RMBG weights in the default self-hosted core.
Browser products have to respect download size, memory pressure, mobile hardware, and first-use latency. A model that wins a benchmark but forces hundreds of megabytes into every visitor's browser can be the worse product decision. The architecture keeps the segmentation layer replaceable so stronger models can be introduced selectively when the quality gain justifies the cost.