Piano2Pop turns simple piano demos into full pop productions without aligned supervision
umpedronosapato · x · 2026-10-28
Researchers unveiled Piano2Pop, a pianoroll-conditioned flow-matching model that expands simple piano demos into full pop arrangements without aligned supervision.
Key technical points:
- Training pairs are pseudo piano-pop pairs generated by MuScriptor and PicoGen, sidestepping scarce real aligned data
- Function-Selective Transcription (FST) uses HTDemucs stem separation and samples four stem-mixture transcriptions (full mix 55%, harmonic / no-bass / piano+bass 15% each)
- Register–Density Augmentation (RDA) transforms the baseline transcription's register and density
- The demo page includes 30-second comparisons against PiCo, MuseControlLite and ACE-Step; authors note PiCo sounds fuller but muddy, while FST adds clearer bass and drums
Code will be released upon paper acceptance.
More from Research
- DeepSeek v4 insider: mHC residual streams may converge — why iHC is the safer pick — teortaxesTex · 2026-10-28
- Qwenputer: a 32-bit RISC-V CPU compiled into Qwen3.5's weights — LatentSpacer · 2026-10-28
- Meta's Loop Scaling Laws: looped MoE matches 2x larger models on reasoning — yacinelearning · 2026-10-28
- Anthropic trial: 77% of analysts echoed their agent's arguments, only 15% credited it — paulnovosad · 2026-10-28
- Generalist Contact Planner + Tracker + Neural Mapping Push Humanoid Mobility Forward — ChongZzZhang · 2026-10-28
- Top econometricians Andrews & Mikusheva publish paper on AI's impact with an AI-written note — paulnovosad · 2026-10-28