This page is the demo

A learned image codec.
Decoded in your browser, byte-for-byte.

wai.neural.int_hyper integer-exact In the standard

What arrives is a wai.neural.int_hyper bitstream: a factorized side stream z and a Gaussian-conditional main stream y. The readout below shows its size in bytes. Your device decodes it with the reference implementation's decoder, compiled to WebAssembly, with no ML runtime and no floating point: factorized z → integer hyper-decoder → integer-σ entropy bucketing → integer synthesis. Because the whole path is integer (the integer-entropy contract), the reconstruction is byte-identical on every conforming sink, and the page re-checks its FNV-1a checksum against the one the native reference recorded for this image.

image — bitstream — model — decoded FNV-1a — native FNV-1a —

Why the whole decode is integer

A float learned codec's entropy decode can desync on a tiny cross-platform difference in the scale σ, so its output reproduces only on the same build. Here every step of the decode is integer, so the checksum above is portable: this browser and a native sink produce the same bytes. The wire carries only the bitstream; the model is the installed codec, fetched once and cached.

The model

A pretrained mean-scale hyperprior image codec from the published learned-compression literature, post-training-quantized to integers. The standard's CI requires the integer decode to stay within 1 dB of the same model's float decode on a pinned test photograph. The page shows the model's size once it has been fetched.

Where the code lives

The capability and its integer decoder are in the standard; the reference engine is Apache-2.0 in the open-standards repository. The image on this page is a demo fixture. The checksum shown is 64-bit FNV-1a; signed receipts hash with BLAKE3.