MIDI-REPA: Improving Text-to-Music Generation with Symbolic Representation Alignment
We introduce MIDI-REPA, a representation-alignment framework that derives MIDI supervision from automatic music transcription (AMT) to align intermediate representations of a flow-matching DiT.
ProofPaper ↗
Key points
- State-of-the-art text-to-music models increasingly rely on scaling model capacity, data, and compute, while still learning musical structure directly from dense audio.
- Building on the 122M-parameter FluxAudio-S baseline from the 2026 ICME Academic Text-To-Music Challenge, we investigate what, where, and when to align.
- Aligning a MuScriptor-derived piano roll with an early DiT layer throughout the full flow trajectory performs the best.
- Our results highlight symbolic supervision as a simple and inexpensive learning signal for improving text-to-music generation.
Sources (1)
- [1]MIDI-REPA: Improving Text-to-Music Generation with Symbolic Representation AlignmentarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 09:42 PM
We introduce MIDI-REPA, a representation-alignment framework that derives MIDI supervision from automatic music transcription (AMT) to align intermediate representations of a flow-matching DiT.
State-of-the-art text-to-music models increasingly rely on scaling model capacity, data, and compute, while still learning musical structure directly from dense audio.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 14, 2026The Design Space of Tri-Modal Masked Diffusion Models
- Oct 14, 2026RepForcing: Representation-Space Initialization for Autoregressive Video Distillation
- Oct 14, 2026CommerceGen-Studio: An End-to-End Framework for Commercial-Grade Main-Image Generation
- Oct 14, 2026Self-Conditioning Enables Deep Sequential Reasoning in Flow Matching
- Oct 14, 2026Glance Forcing: Making Bidirectional Diffusion Autoregressive at a Glance
- Oct 13, 2026EqC: Equilibrium Energy Control for Introspective Lifelong Robot Learning