BiRefNet Lite (FP16, 98 MB) in onnxruntime-web on a single WASM thread: streaming load, memory settings, low-RAM gating

[Mehmet (xenitV1)]

Sharing a small production use of onnxruntime-web in case the details help someone doing the same: a background remover that runs BiRefNet Lite entirely in the browser, on the plain WASM backend.

Live: https://sticraft.com/background-eraser (the "detailed cutout" button)

Setup

• onnxruntime-web 1.27.0, onnxruntime-web/wasm import, executionProviders: ["wasm"], numThreads = 1. The site isn't cross-origin isolated, so WASM threads aren't available; one thread in a dedicated worker keeps the page responsive.
• Model: BiRefNet Lite 512, FP16 ONNX export (98,484,532 bytes). Input 1x3x512x512 with ImageNet normalization; the single output is passed through a sigmoid to get the alpha mask, then scaled back to the photo.
• Weights and the ORT .mjs/.wasm files are served from the same origin, so a strict CSP connect-src 'self' holds and no image or model request leaves the site. The build pins the model by SHA-256.

Things that mattered for memory

• The model is streamed into one preallocated Uint8Array of the exact size (which also gives an honest download progress bar), passed to InferenceSession.create, and the reference is dropped right after the session is created.
• enableCpuMemArena: false and enableMemPattern: false: there is exactly one inference per session, so there is nothing for them to reuse.
• A fresh, short-lived worker per detailed cutout. Terminating it afterwards releases the whole WASM heap, and the versioned model file stays in the normal browser cache for the next run. Cancel and a 5-minute timeout also just terminate the worker.
• Devices that report navigator.deviceMemory <= 4, or phone-sized touch viewports when that hint is missing, don't get the detailed option at all; they keep a small U²-Net (u2netp) first pass. Total first-use download for the detailed path is about 112 MB including the runtime, and the UI says so before starting.

Open question for anyone who has tried it: has WebGPU with FP16 BiRefNet been stable for you on integrated GPUs? I stayed on WASM because I couldn't test enough hardware.

Disclosure: I'm the developer of the site, and this post was drafted with an AI assistant from my project notes and code.