kickvn/kim-inst-kick-mlx

Kim Inst (MDX-Net) - KICK Core ML / MLX

The instrumental separation model Kim Inst (UVR MDX-Net, Kim_Inst.onnx) converted by KICK for Apple Silicon Macs, in two separate packages: Core ML (the network runs on the Apple Neural Engine, fp16) and MLX (GPU). KICK Studio on macOS uses the Core ML package to separate the backing track (beat) from a song.

Speed and accuracy

Full song of 221.6 s on a MacBook Air M4 (8-core GPU, fanless):

Path	Time
Core ML on the Neural Engine (network), overlapped with MLX STFT/iSTFT on the GPU	4.8 s
Core ML on the GPU	17 s
MLX on the GPU, fp32	27 s

• PyTorch rebuild of the network vs the original ONNX: mean relative difference 8e-5.
• Neural Engine (fp16) vs fp32 output: SNR 60.4 dB.
• MLX package (fp16-stored weights) vs fp32: SNR 67.1 dB.

Files

File	Size	Contents
kim-inst-kick-coreml.zip	~30 MB	folder kim-inst-coreml/: mdx.mlpackage + mdx_config.json
kim-inst-kick-mlx.zip	~30 MB	folder kim-inst-mlx/: mdx.safetensors + mdx_config.json

• mdx.mlpackage - Core ML ML Program, fp16. Input spec float32 [1, 4, 3072, 256], output out same shape (channels: left real, left imag, right real, right imag; 3072 frequency bins x 256 frames).
• mdx.safetensors - the same weights for MLX (BatchNorm folded into scale/shift, conv weights [out, kh, kw, in]), stored in fp16, computed in fp32.
• mdx_config.json - architecture and STFT settings.

STFT around the network: n_fft 7680, hop 1024, Hann window, center/reflect padding, first 3072 bins kept, the 3 lowest bins set to 0 before the network. Chunks of 261,120 samples (256 frames) with 25 % overlap and a Hanning cross-fade, as in UVR.

Credits

• Model: Kim Inst by Kimberley Jensen, distributed with Ultimate Vocal Remover (UVR, Anjok07 and aufr33): original Kim_Inst.onnx from TRvlvr/model_repo.
• Architecture: MDX-Net / ConvTDFNet (TFC-TDF U-Net), KUIELab.
• MLX weight conversion: based on ssmall256/mlx-audio-separator (MIT).

License

The upstream weights were released without an explicit license. This repository does not grant any additional rights; the terms of the original weights apply. If you are the author and want this removed, open a discussion.

Core ML / MLX conversion by KICK.