Text-to-music models typically learn musical structure directly from dense audio, even though melody, rhythm, harmony, and structure can be represented compactly in symbolic form. MIDI-REPA, a representation-alignment framework, derives MIDI supervision from automatic music transcription and uses it to align intermediate representations of a flow-matching diffusion transformer.
The authors build on the 122M-parameter FluxAudio-S baseline from the 2026 ICME Academic Text-To-Music Challenge. They systematically test what to align, where to align, and when to align. The best configuration aligns a MuScriptor-derived piano roll with an early DiT layer throughout the full flow trajectory.
Without adding training data or new techniques, MIDI-REPA remains competitive on the ATTM objective metrics while remaining trainable on a single RTX 3090 in about 31 hours. The authors highlight symbolic supervision as a simple and inexpensive learning signal for improving text-to-music generation. They also release their transcription pipeline, JamendoMIDI, training code, and audio samples.