magenta-realtime-2-onnx

Magenta RealTime 2 - ONNX conversion (mirror)

ONNX conversion of Google DeepMind's Magenta RealTime 2 weights, hosted for the Magenta RT 2 VST3 / Standalone plugin's first-run model downloader.

• Model: Magenta RealTime 2 (c) Google DeepMind
• License: Creative Commons Attribution 4.0 International (CC-BY-4.0) - https://creativecommons.org/licenses/by/4.0/legalcode
• Changes: these files are an ONNX conversion of the original JAX/Flax weights (the upstream release ships no ONNX). Converted for on-device ONNX Runtime inference.
• Responsible use: use these materials responsibly; do not generate content that infringes the rights of others. Google claims no rights in outputs you generate.

Layout

Two variants, each a self-contained directory:

mrt2_small/  manifest.json  NOTICE  LICENSE-CC-BY-4.0.txt
             main_trans.onnx(+.data)  noise.onnx  main_dec.onnx(+.data)  encoder.onnx(+.data)
             cpp_duo_violin/  index.json  state_*.bin  ...
mrt2_base/   (same file set, base-sized transformer)

manifest.json is schema-1 (relative-path name + size + SHA-256). The plugin fetches <repo>/resolve/main/<variant>/manifest.json and verifies every file by SHA-256 on download.

Note on mrt2_base

mrt2_base/main_trans.onnx is over 2 GB. TensorRT's ONNX parser cannot ingest a

2 GB graph directly, so a downloaded mrt2_base still requires a machine-local TensorRT engine build on first load. mrt2_small loads directly (download, then play).