Model and limits
This implementation uses the MIT-licensed studioludens/birefnet-lite-512 browser ONNX export with Transformers.js. Its 512×512 inference resolution is practical for browsers but may soften very fine edges on large source images.
Device behavior
WebGPU is preferred when available; WASM is the fallback. Actual processing time is measured and shown after each run instead of claiming a universal speed. The model is loaded only on this tool page when processing starts.