notabilia/uvr5-models

UVR5 MDX-Net Models — Backup Repository

This repository is a backup and redistribution mirror of select pre-trained ONNX models from the Ultimate Vocal Remover (UVR) project. The models are hosted here unaltered for reliability and availability purposes. This repository is not affiliated with the original authors.

Purpose

The canonical source for these models is the TRvlvr/model_repo GitHub release archive. This mirror exists to provide a stable, versioned download location independent of GitHub release availability.

Redistribution is permitted under the MIT License, which explicitly grants the right to use, copy, and distribute the models provided that the original copyright notice and license text are preserved. Each model in this repository is accompanied by a .LICENSE file containing the full attribution and license terms.

Models

File	Size	Description	Original Source
UVR-MDX-NET-Inst_HQ_3.onnx	64 MB	High-quality instrumental/vocal separation	TRvlvr/model_repo
UVR_MDXNET_Main.onnx	64 MB	Main MDX-Net vocal separation model	TRvlvr/model_repo
UVR_MDXNET_KARA_2.onnx	50 MB	Karaoke stem separation (v2)	TRvlvr/model_repo
UVR_MDXNET_KARA.onnx	28 MB	Karaoke stem separation	TRvlvr/model_repo
UVR_MDXNET_3_9662.onnx	28 MB	MDX-Net v3 (96.62 SDR)	TRvlvr/model_repo

Architecture

All models use the KUIELab-MDX-Net architecture and are exported in ONNX format for cross-platform inference.

• Paper: KUIELab-MDX-Net: A Two-Stream Neural Network for Music Demixing
• Architecture authors: Woosung Choi, Minseok Kim, Jaehwa Chung, Daewon Lee, Soonyoung Jung (KUIELab)

Credits

All models were trained and published by the Ultimate Vocal Remover team:

• UVR Project: Anjok07 (DilanBoskan), KimberleyJSN, and the UVR development community
• UVR Repository: https://github.com/Anjok07/ultimatevocalremovergui
• Model Distribution: https://github.com/TRvlvr/model_repo

If you use these models in your work, please credit the original UVR authors.

License

MIT — see the .LICENSE file accompanying each model for the full license text and attribution notice.