A learned image codec.
Decoded in your browser, byte-for-byte.
What arrived over the wire is a wai.neural.int_hyper bitstream — a factorized
side-stream z and a Gaussian-conditional main stream y, about
~2.1 KB. Your device decodes it with the same pure-i64 Rust native sinks run,
compiled to wasm32: no onnxruntime-web, no WebGPU, no float anywhere. The whole
path is integer — factorized z → integer hyper-decoder → integer-σ entropy
bucketing → integer synthesis. Because of that (the integer-entropy contract) the
reconstruction is byte-identical on every machine, so the page re-checks its FNV-1a against
the hash the native reference produced. A learned codec, verified client-side, no
server. (The model is a real pretrained codec post-training-quantized to integer, ~0.6 dB
off the float baseline; ~23 MB, cached after first load. The decode runs a couple of
seconds of pure-integer math, stated plainly.)
Why this is the moat
A float learned-image decode desyncs on ~1e-5 of cross-platform drift in the
scale σ — reproducible only on the same build, never
receiptable. int_hyper is integer all the way down, so the hash above is portable:
this browser, a native sink, and the conformance corpus all produce the same bytes. The wire
carried a ~2.1 KB bitstream; your device produced the exact image, and proved it. Reference
engine + conformance corpus in the open-standards repo, Apache-2.0.