When you want to unmix a song but don't want to pip install torch

What it does

Demucs-Gui packages Meta’s Demucs music-separation model into a clickable desktop application for Windows, macOS, and Linux. It takes a mixed audio file and uses machine learning to isolate drums, bass, vocals, and other stems—no Python scripts or virtual environments required. Because the underlying engine depends on PyTorch, the bundled binaries are hefty, but they are meant to run standalone.

The interesting bit

The project’s entire value is subtraction: it strips away terminal friction from a capable open-source AI tool. Despite being essentially a GUI layer, it supports an unusually broad set of hardware accelerators—NVIDIA GPUs, Apple Silicon via MPS, Intel Arc and Iris Xe, and AMD cards on Linux through ROCm. The author is also refreshingly candid that Intel integrated graphics can be slower than CPU, which suggests the docs are written by a human, not a marketing bot.

Key highlights

• Ships as standalone binaries; end users do not need to install Python or manage dependencies
• Supports FFmpeg for audio ingestion when it is present on the system
• GPU acceleration spans NVIDIA, Apple MPS, Intel, and AMD (Linux only)
• Some release binaries exceed GitHub upload limits and are mirrored to FossHub
• Explicitly targets users with zero coding experience

Caveats

• macOS Gatekeeper aggressively blocks the app; the README includes detailed bypass instructions for both pre- and post-15.0 systems
• PyTorch is pinned to version 2.0.x because the author notes that 2.1.0 and higher “cause errors sometimes”
• Intel integrated graphics are officially supported but actively discouraged, since they can underperform compared to pure CPU separation

Verdict

Grab this if you are a musician, producer, or DJ who needs clean stems and does not want to learn conda. If you are already comfortable with Python and PyTorch, you are probably better off running Demucs directly and skipping the middleman.