noupload — background removal model benchmark

MODNet (fast tier) vs BiRefNet Lite (quality tier). Same 4 images, same machine (M4). Matte composited on checkerboard.

Model facts

	MODNet fp16	BiRefNet Lite fp32
Download (one-time)	12.4 MB	213.6 MB
Input	512×512	1024×1024
Load	0.63 s	2.1 s
Per image (avg of 3)	1.48 s — in-browser, Chrome CPU (wasm)	11.1 s — native onnxruntime CPU (M4). Browser wasm build: over 4 min/image single-threaded, needs WebGPU to be usable
Edge quality	good on portraits, weak on fur and general objects	excellent — hair, fur, complex edges
Scope	portraits/humans only	general: objects, animals, products
License	Apache 2.0	MIT

Verdict

MODNet is ~7.5x lighter and ~7.5x faster, but portraits only. BiRefNet Lite is the quality king for everything (hair, fur, objects) and it is MIT licensed — no traps. Recommendation for noupload: two tiers — MODNet as the instant preview, BiRefNet Lite as the export pass, shipped with WebGPU only. Measured without WebGPU (plain CPU wasm), BiRefNet took over 4 minutes per image in a tab, so WebGPU is not optional for this model.

Timing environment: same M4 machine. MODNet: Chrome browser, wasm CPU. BiRefNet: native onnxruntime CPU (the browser wasm path was not viable — that itself is the finding). First-run model downloads are cached by the browser after initial use.