Stem Separator: One-Click AI Stem Enhancement with Apollo

The Stem Separator just got a major quality upgrade: every stem can now be AI-enhanced with one click. Source separation inevitably leaves artifacts behind — spectral "musical noise", smeared transients, and lost high-frequency shimmer. The new ✨ Enhance button repairs exactly that.

What's new

• Per-stem Enhance button — each stem card (vocals, drums, bass, guitar, piano, other, instrumental) now has a ✨ Enhance button that runs the stem through Apollo, a state-of-the-art music restoration model from Tsinghua University / Tencent AI Lab (JusperLee/Apollo on HuggingFace, CC BY-SA 4.0).
• ✨ Enhance All — one click queues AI restoration for every stem in the task. Jobs run sequentially on the GPU alongside separation, with live per-stem status (queued → enhancing → done).
• A/B comparison — once a stem is enhanced, an A/B toggle appears on the card: swap playback between the original and enhanced versions instantly, with the waveform redrawn for each.
• Enhanced downloads — grab the enhanced file directly (⬇ enhanced) or get everything via Download All — the ZIP now includes both original and enhanced versions.

How it works

Apollo is a band-split GAN trained on MUSDB18-HQ and MoisesDB — real multi-instrument music stems. It works in the frequency domain, splitting the spectrogram into sub-bands and regenerating the mid/high-frequency detail that separation masking tends to erase, while suppressing the broadband noise artifacts that leak through. Processing a full 4-minute stem takes under a minute on the RTX 2070.

Tip: enhancement is restorative, not regenerative — use the A/B toggle to judge each stem. For vocals and acoustic material the difference is often most audible in the "air" above 10kHz.