1. Introduction
MuLaCover is a cover-song generation model that conditions a pretrained text-to-song foundation model on a symbolic lead sheet. The user can freely combine reference music, lyrics, and style to produce a new cover. We support two task families:
- Full-Song Cover. Re-render an entire song under a new style. The reference can be combined with either the original lyrics or with new lyrics provided by the user.
- Partial-Reference Song Cover. Only a portion of the reference is observed; the rest is freely generated. Three configurations: continuation (only the song's beginning, up to the first chorus, is visible), outpainting (only the first chorus is visible), inpainting (everything except the first chorus is visible).
Below we showcase MuLaCover under each configuration on a single demonstration song. Baselines are excluded here; quantitative comparisons appear in Sections 2 and 3.
1.1 Full-Song Cover Demo
[Loading...]
[Loading...]
1.2 Partial-Reference Cover Demo
[Loading...]
2. Full-Song Cover Task
We compare MuLaCover with three baselines: Suno v5.5, ACE-Step 1.5, and SongEcho. Eight cases in total: four use the song's original lyrics under a new style; the other four use new lyrics under a new style. Only the first case in each subsection is expanded by default — click any collapsed case header to reveal it.
2.1 Original Lyrics + New Style (Cases 1–4)
2.2 New Lyrics + New Style (Cases 5–8)
3. Partial-Reference Cover Task
We compare MuLaCover with two baselines: ACE-Step 1.5 and SongEcho. Suno v5.5 is excluded. Six cases — two for each configuration. The first case in each configuration is expanded by default; click a collapsed header to reveal the rest.