YuE2 Studio runs M-A-P's music model locally on 6GB VRAM with editable scores and LoRA training
wzwowzw0002 · reddit · 2026-10-21
A Reddit user highlights YuE2-Studio, an open-source standalone Windows desktop app that runs M-A-P's YuE2 model locally on any GPU with 6GB+ VRAM (no Python setup) to generate full songs from text prompts and lyrics.
Key features:
- Editable sheet music: generates ABC-notation scores first; edit notes, key, chords, or tempo and re-render
- Exact replays & checkpoints: saves audio codes and parameters for bit-for-bit re-renders or tweaking specific checkpoint tokens
- Hugging Face LoRAs: built-in browser to search, download, and apply LoRA/LoKr weights
- On-card LoRA training: a trainer tab to train custom LoRAs on your own audio datasets directly on your GPU
- Audio & stem tools: tracks up to 10 minutes, GPU stem separation (HT-Demucs), Audio-to-MIDI, piano roll editor, and word-level karaoke timestamps
The author confirms it works well standalone and can also be driven via CLI with hermes.
More from Multimodal
- Mirelo SFX 2.0 auto-designs full soundtracks for AI videos, no full regeneration needed — iamfakhrealam · 2026-10-21
- NVIDIA demos interactive AI surgical simulator running on a single RTX PRO 6000 GPU — Scobleizer · 2026-10-21
- Topaz joins Adobe; Starlight Fast upscales 720p AI video to 4K/60fps — Mr_AllenT · 2026-10-21
- FLUX 3 Klein on RTX 3090: int8 Doubles Speed, but FLUX.2 klein 4B Wins on Quality — strata2signal · 2026-10-21
- EvaOmni: an evidence-grounded benchmark for audio-visual reasoning in omnimodal LLMs — VisionXLab · 2026-10-21
- WinARDet: compressed window autoregression stops box repetition loops in dense grounding — NankaiUniversity · 2026-10-21