WebGPU ran in-browser cutout 6x faster but drew the wrong mask

A Koutuxia developer found that the WebGPU backend in ONNX Runtime Web 1.27 produced wrong logits for a quantized BiRefNet model: about 75% of the frame fell into the semi-transparent band versus 0.8% on WASM. WebGPU was disabled in production on September 13, and switching to fp16 weights in WASM cut inference from 16.8 s to 6.95 s.

• On WebGPU the int8 model put 75.1% of the frame in the soft band versus 0.8% on WASM
• fp32 on WebGPU matched WASM, pointing to the int8 DequantizeLinear path
• On seven AIM-500 objects WASM beat WebGPU: median SAD 1.10 versus 6.15
• fp16 weights in WASM cut inference from 16.8 s to 6.95 s with alpha difference around 1e-5