Extending stem separation to guitar and piano is harder than the standard four-stem split

Standard Demucs separates four stems: drums, bass, vocals, other. The htdemucs_6s model adds guitar and piano. At the time of the v4 release, guitar quality is acceptable but the piano stem shows significant bleed and artefacts. This reflects a fundamental challenge: piano and guitar share harmonic spectral regions with each other and with ‘other’ instruments, and isolated piano/guitar stems are scarce in freely licensed multi-track datasets. The ‘other’ category absorbs most remaining harmonic content in the four-stem model, but splitting it requires much more training data with clean guitar and piano stems.

Examples

# Try the 6-stem model:
demucs -n htdemucs_6s mytrack.mp3
# Outputs: drums.wav, bass.wav, vocals.wav, other.wav, guitar.wav, piano.wav
# Warning: piano.wav may have significant bleed

Assessment

Why is separating piano harder than separating drums from a neural model perspective? What properties of drums make them easier to isolate?