swwallowws/coming-undone-browser-models

Coming Undone: browser models

Model files for the "In your browser" engine of Coming Undone, which separates a song into stems and transcribes them to MIDI entirely in the visitor's browser (onnxruntime-web, WebGPU with a wasm fallback). The page reads models.json from this repo and downloads the files it lists once, into the browser's Cache Storage. Nothing is uploaded: the audio never leaves the machine.

Files

File	Size	What
adt_encoder.int8.onnx	35.7 MB	ADT_STR drum transcription, encoder: log-mel (B, 246, 128) in, each decoder layer's cross-attention K/V out
adt_decoder.int8.onnx	37.2 MB	ADT_STR drum transcription, decoder: token prefix + cross K/V in, next-token logits out (greedy decode runs in JS)
models.json	1 KB	the manifest the page reads: file names, sizes, sha256, licences, and the pinned htdemucs URL

Both ADT_STR graphs store their weights as int8 (per channel, weights only). The page turns them back into float weights before it creates the session, because onnxruntime-web 1.30's WebGPU backend computes DequantizeLinear feeding MatMul wrong. Against the original PyTorch model on the same drum stem the int8 graphs give the same hits (onset F1 1.000 on a 20 s test clip, 97/97 hits).

Where they come from

• ADT_STR (Melucci, Merialdo, Akama 2026), variant setting-tau-0.8 from Pierfrancesco/adt-str at revision a33c5c6b191a4ca1e0f6dc22140947485eb36ce8. Code: github.com/pier-maker92/ADT_STR. Exported to ONNX (attention written out by hand over the same weights) and quantized to int8 by Coming Undone's browser/models/export_adt.py and browser/models/quantize.mjs.

Not in this repo:

• htdemucs separation (Défossez et al., Meta, facebookresearch/demucs, MIT): the page loads the ONNX export by timcsy straight from timcsy/demucs-web-onnx, pinned to revision 92e33df61cfc9eb820272aaa62d2ef6dcf4d950d (180.5 MB). That repo states no licence of its own.
• basic-pitch (Spotify, Apache-2.0): its 0.9 MB TF.js model ships with the page.
• MuScriptor small (Mirelo and Kyutai): its weights are gated and CC BY-NC 4.0, so they are not shared here. MuScriptor runs in Coming Undone's Online and This computer engines instead.

Licence

The ADT_STR weights are licensed under Creative Commons Attribution-ShareAlike 4.0 International by their authors: Melucci, Merialdo, Akama, "ADT_STR" (2026), github.com/pier-maker92/ADT_STR.

The files here are an adaptation of those weights (exported to ONNX and quantized to int8). They are shared under the same licence, CC BY-SA 4.0: you may use them, including commercially, as long as you credit the original authors, say that the weights were changed, and share your own adaptations under CC BY-SA 4.0 too. The quantized weights stay CC BY-SA 4.0 whatever format they are converted to.