This page is the demo

A 3D cloud you can put in a receipt.
The same Gaussians, everywhere.

wai.splat.int_codec splat-equivalence In the standard

A Gaussian splat is a tensor of per-Gaussian attributes: position, scale, rotation, colour and opacity. wai.splat.int_codec is a learned codec for that tensor whose synthesis is a stack of integer convolutions and integer IGDN over a factorized-rANS latent. This page runs the reference implementation's decoder, compiled to WebAssembly, with no ML runtime and no floating point. So the decoded attribute tensor is byte-identical on every conforming sink, and the page re-checks its checksum against the native reference before the sink's rasterizer draws it. That is splat-equivalence: the attributes are the contract; the pixels are the sink's.

decoding…
gaussians — attrs — decode — attribute hash (FNV-1a) —

Why the whole decode is integer

Before the cloud was drawn, the page hashed the decoded attribute tensor and checked it against the checksum the native reference recorded; the badge shows the result. The transposed convolutions are integer accumulation with an arithmetic-shift requantize, IGDN uses an integer square root, and the latent comes off an integer rANS decoder. The rasterizer that presents the cloud stays the sink's choice: WAI pins the attributes, not the pixels.

The model

A small codec trained on a synthetic, procedurally generated cloud, then post-training-quantized to integers; the page shows its size once fetched (— KiB). It demonstrates the integer decode.

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. Its test vectors live with the reference implementation (wai-rs/tests/vectors/int_splat). The cloud on this page is a demo fixture. The attribute hash shown is 64-bit FNV-1a.