MuLaCover open-sources a controllable model that rebuilds any song into an AI cover
solyarisoftware · x · 2026-09-17
The HeartMuLa team released MuLaCover on Hugging Face: a controllable AI cover-song and music-remix model with weights freely downloadable (CC-BY-NC-4.0, paper arXiv:2601.10547).
- Flexible inputs: drop in any track or MIDI, swap the lyrics, and describe a vibe to rebuild the song from the ground up
- Two control paths: lyrics plus a structured one-line style prompt (topic/genre/instrument/mood fields), or melody/chord MIDI generation
- Controllable dimensions: remixes genre, instrumentation, topic, and mood; audio is first converted to symbolic melody/harmony/drums
- Setup: requires sibling checkpoints (HeartCodec, Qwen3-Embedding-0.6B, YourMT3, ChordNet), inference via the mulacover pipeline in bfloat16
More from Multimodal
- First hands-on with SpAItial Echo: 3D world models from a single image — jonstephens85 · 2026-09-17
- Seedance 2.5 kung fu demo paired with Suno v6 generated soundtrack — TinfoilTricorn · 2026-09-17
- Luma details Ichor pipeline: Seedance 2.5 edit mode plus LED walls for performance — mrjonfinger · 2026-09-17
- World Labs Atlas Beta Rolls Out to Users With API Ideas From 3D Reconstruction to Graybox-to-World — theworldlabs · 2026-09-17
- AI-Generated Animated Short 'The One Who Walks at Night' Prologue Released — Outside-Objective828 · 2026-09-17
- Indie filmmaker spends 1.5 months crafting AI sci-fi short FIDELITY for Higgsfield festival — eraiozei · 2026-09-17