Filling MIDI Velocity using U-Net Image Colorizer

Zhanhong He, David Cooper, Defeng Huang, Roberto Togneri · 2025

Read Filling MIDI Velocity using U-Net Image Colorizer by Zhanhong He, David Cooper, Defeng Huang, Roberto Togneri (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-07751.txt.

Description

Zhanhong He and co-authors published this music technology study in 2025. MIDI files store music as note data, and each note carries a velocity value that sets its loudness. Files made in software often leave velocity flat, so the music sounds mechanical. Predicting realistic velocities can restore a human feel. The team borrowed U-Net, a network famous for coloring images, and treated MIDI data as pictures to fill in.

Their design adds window attention and a custom loss suited to the sparse look of MIDI images. Experiments were limited to piano music, since suitable datasets cover little else. On the MAESTRO v3 and SMD piano datasets, the method beat earlier approaches on numeric scores. It also won in listening tests, where judges rated the filled-in performances. The authors argue that image tools, reshaped carefully, can serve music tasks that seem far from vision.

Catalog record

CREATORZhanhong He, David Cooper, Defeng Huang, Roberto Togneri
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE21 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-07751

Browse the full library · About the Tower of Babel