Kirexa MDX-Net — LiteRT models

.tflite builds of the MDX-Net music source-separation models used on-device by Kirexa (Android), converted from the original .onnx via onnx2torch → litert_torch.

The app uses the fp32 files. They run on the LiteRT GPU delegate, which executes them in fp16 internally (precisionLossAllowed) — so we get GPU fp16 speed without baking float16 into the file. (A float16-quantized .tflite carries explicit DEQUANTIZE nodes that shatter GPU graph partitioning and run an order of magnitude slower on-device; fp32 has none.) On devices without a usable GPU the same fp32 file runs on the XNNPACK CPU backend.

Files (use these — fp32)

Input tensor [1, 4, dim_f, dim_t] (NCHW: 2 channels × {real, imag}), float32. The host app computes the STFT, runs the model, and applies the iSTFT.

The *_fp16.tflite files are deprecated (half the size, but the DEQUANTIZE nodes cripple the GPU delegate). Kept only for reference — do not use.

Attribution

Converted re-hosts of third-party weights — credit to the original authors:

• UVR-MDX-NET-Voc_FT (vocals) — Ultimate Vocal Remover (Anjok07).
• kuielab_a_{drums,bass,other} — kuielab/mdx-net, Kim et al., Sony Music Demixing (MDX) Challenge 2021.

License

MIT, inherited from the upstream MDX-Net / UVR projects. The .tflite conversion adds no new restrictions.