Demucs automatically rescales output stems to prevent clipping but this breaks relative stem loudness

When Demucs separates a mix, the predicted stems may sum to a value exceeding 0 dBFS, causing clipping if saved naively. By default Demucs rescales each output stem independently to avoid clipping. This preserves each stem’s waveform shape but destroys relative loudness relationships — drums may be at the same peak level as a quiet guitar line. If you need stems that reconstruct the original mix when summed, use --clip-mode clamp (hard clip) or reduce input level. For sampling purposes, the default rescaling is usually preferable because it maximises the dynamic range of each stem individually.

Examples

# Default: auto-rescale per stem (breaks relative levels)
demucs mytrack.mp3

# Hard clip instead (preserves relative levels, risks clipping)
demucs --clip-mode clamp mytrack.mp3

Assessment

When would you choose --clip-mode clamp over the default rescaling? Describe a use-case where the default behaviour would cause problems.