Open Source
Explore the latest AI open-source projects from GitHub and HuggingFace.
Explore the latest AI open-source projects from GitHub and HuggingFace.
Remiqora is a local, GPU-powered music generation and production studio that puts two independent engines, ACE-Step 1.5 and YuE2-3B, behind a single Vue 3 interface with a FastAPI backend. The README describes it as an orchestrator rather than a generator: the project does not vendor the third-party engines, only small patches on top of them, and a mutually exclusive orchestrator starts one model when you pick it in the header and stops the other, since the two typically cannot share a consumer GPU. ACE-Step handles text and style-tag generation, covers, section repainting, and extracting or adding instrument parts on a reference track. YuE2-3B generates full-length tracks with an optional chain-of-thought plan, and SheetSage2 can pull a reference melody into ABC notation as its input. A LoRA training pipeline for ACE-Step runs in the browser from dataset upload through auto-labeling, preprocessing, training and export. Demucs splits a track into four stems (vocals, drums, bass, other), and MuScriptor transcribes audio to MIDI, which requires YuE2 to be the active model. A built-in multitrack DAW offers clip splitting, fades, loop regions, BPM detection with time-stretch, an eight-effect rack per channel and on the master, up to 30 undo steps, and WAV or MP3 export. Every track lives in a shared SQLite library. The repository is MIT-licensed, but generated audio follows the license of the model that made it: the README states YuE2-3B weights are CC BY-NC 4.0, while ACE-Step 1.5 code and weights are MIT. The project is marked as actively in development with expected breaking changes, and an experimental Electron desktop app is offered for Windows and Apple Silicon macOS, with roughly 30 GB of first-launch downloads.