RhythmForge

Stem separation quality

• Server AI Hybrid ★★★★★ — Demucs tier-1 (htdemucs_ft) + WHAMR, 14–20+ dB SDR (dedicated GPU worker).
• Server AI ★★★½ — Hugging Face ML, ~8–14 dB SDR.
• Client Enhanced ★★★½ — browser Wiener masking, ~6–10 dB SDR, orthogonal bleed reduction.
• HF Fallback ★★ — server AI failed; Client Enhanced ran instead.
• Simulated ★ — demo stems when no backend is configured.

Best results: deploy Demucs worker (GPU), or configure HF_API_KEY for Server AI. Upload audio directly for Client Enhanced. Enable Prefer Local Processing to skip server mock stems on URL imports.

Auto remix

• Genre presets — fastest one-click remix (Cyberpunk, Lo-Fi, Dance, etc.).
• Mood prompt — describe style/energy; engine applies stem-aware EQ, ducking, dynamics.
• Decisions panel — collapsible summary of what was applied (after auto mix).
• Playback automation — subtle volume/EQ/reverb changes during ▶ Preview only (not in export).

Server AI setup (deployers)

1. Create a Hugging Face account and generate an Inference API token.
2. Add HF_API_KEY to your Vercel/host environment variables.
3. Redeploy — the Server AI status banner will show "Configured".
4. If HF fails (quota/timeout), Client Enhanced runs automatically.

Export formats

• WAV — lossless offline mix (best quality).
• MP3 — browser-encoded via lamejs; 128/192 kbps.
• WebM — live capture while transport plays.

💡 On mobile, enable Perf mode or Low Quality for smoother visualizer and playback. Try Aurora Lanes for per-stem reactivity.

💡 Connect your Xaman wallet to mint NFTs and earn royalties.