KUIELab-MDX-Net

KUIELab-MDX-Net won 2nd place on Leaderboard A and 3rd on Leaderboard B of the Sony Music Demixing Challenge at ISMIR 2021, and it did it by refusing to choose sides: one branch works in the time-frequency domain (a lean U-Net with time-distributed fully-connected layers), the other works directly on the waveform (Demucs-style), and the model blends both streams' outputs. The design came from deliberately trading parameter-heavy dense blocks for efficient stacked convolutions plus TDF layers, so it runs well without a datacenter GPU.\n\nIts bigger legacy is cultural: the UVR (Ultimate Vocal Remover) ecosystem's legendary 'MDX-Net' vocal models — the ones that made free vocal isolation dramatically better than Demucs and Spleeter for most users — are this architecture. The code is MIT-licensed and the challenge submission repos ship pretrained models.\n\nIt is 2021-era now, and Transformer-based separators have passed it on raw SDR, but for vocals-vs-instrumental splitting it remains a respected, well-understood baseline with a huge installed base.

• Challenge-verified: 2nd place Leaderboard A, 3rd place Leaderboard B, Music Demixing Challenge @ ISMIR 2021\n- The architecture behind UVR's default MDX-Net vocal models, widely reported by users to outperform Demucs and Spleeter on vocal isolation\n- 2021-era: superseded on MUSDB18 SDR by Transformer-based models (BS-RoFormer family)