YuE2 Studio

A Windows desktop app that runs YuE2, m-a-p's song-generation model, on the user's own GPU: full songs with vocals, an editable score, and no Python or Node.js to install. One executable, NSIS installer with auto-update or a portable archive, offline once the model is downloaded. Read from the repo README at read time (2026-09-24); via Telegram. Runner-up spoke: music-tech-wiki, which owns synths, effect plugins and DAW software. This is a model-consumption app around a neural song-generation model, which is this spoke's territory, not music-tech's.

What it is

A Rust/React app in a Tauri window, calling yue2.cpp (ServeurpersoCom/yue2.cpp), a native C++/CUDA port of YuE2, through a local Rust/Axum service on port 8791 plus a yue-server process on port 18087. It grew out of the author's earlier MiniMax Music3 Studio. The studio and yue2.cpp are both MIT; the models it runs (YuE2-3B, its VAE, and the optional sheetsage2 transcriber) are CC BY-NC 4.0, non-commercial only.

Features

Full songs, up to six minutes, from a style prompt plus lyrics.
Editable score: the engraved ABC melody and chords (notes, tempo, key) can be changed, then regenerated. Other modes skip the score (melody-only, no score) or plan the score without audio.
Covers: sheetsage2 transcribes an existing recording's melody, and YuE2 sings it in a new style. This is the first place in the wiki sheetsage2 is actually put to use rather than described as a standalone model. The lyrics have to match the source's syllables and stresses, and transcription errors carry straight into the cover.
Exact replay: every request and its audio codes are saved, so a track re-renders bit-for-bit, or with different steps, seeds or output formats, later.
Karaoke (word-level LRC via Parakeet or Whisper), six-stem separation (HT-Demucs), LoRA training and a LoRA catalogue (via a third-party trainer, on 5–20 songs), a writing assistant on a local Gemma model or OpenRouter, noise reduction and VST3 mastering chains, MP3/WAV output, ID3 tagging and playlists. UI in English, Russian, Chinese, Japanese and Korean.

Hardware and model sets

Minimum 6 GB VRAM, GTX 900-series or newer NVIDIA (CUDA 13 needs Turing+/driver 580+; CUDA 12 covers Maxwell through Volta or older drivers); AMD/Intel support is experimental via Vulkan, and the README states vocals came out unintelligible on AMD Radeon integrated graphics. CPU fallback exists and is "much slower." Four quantized model sets trade VRAM for size: BF16 (12 GB+, 9.7 GB download), Q8_0 (8 GB+, 4.9 GB), Q6_K (7 GB+, 4.0 GB), Q5_K_M (5.5 GB+, 3.6 GB), all against Serveurperso/YuE2-GGUF pinned to revision 64b030e. Training a LoRA needs an RTX 30-series or newer with 11 GB VRAM.

One self-reported speed figure, flagged as such: a 3:38 song at Q8_0 renders in about 46 seconds on an RTX 4090. That's the maker's own number, not an independent measurement, and the only performance claim in the README.

Authors and repo

timoncool on GitHub (ilya-timonin, posting as "Nerual Dreming") and Нейро-Софт; 104 stars, 8 forks, 102 commits, 0 open issues or PRs at read time. Credits M-A-P for YuE2, Serveurperso for the yue2.cpp port and GGUF quantizations, plus scragnog, sergree (matchering), jeankassio, the LoRA authors and LAME.

What it adds here

A third instance of the consumption-layer pattern vocalinux and abogen established: an application that supplies everything around an open model and contributes no model of its own, now on the music-generation branch instead of STT or TTS. It is also the first page here to show sheetsage2 doing real work rather than sitting as a thin, freshly-paged model card: covers depend on its transcription being good enough to carry lyrics into a new performance, which is exactly the tradeoff that page's melody-F1 number (82.51%, one note in five wrong) already flagged.

ilya-timonin's own GitHub profile lists several more "Studio" wrapper apps built the same way around other open audio models (Qwen3-TTS, HiggsAudio, MiniMax Music3, DotsTTS, VoxCPM2, ACE-Step), none of them ingested here yet. The pattern is one developer's repeated move, not a one-off.

Tier T1: the load-bearing claims are the app's own facts (architecture, feature list, licence, platform and hardware requirements, repository metadata), which is what the artifact-facts branch of the tier rule grades T1 for a first-party repo. The one self-run benchmark (the 46-second render) is flagged above rather than folded into the tier.

yue2 · multimodal-art-projection · sheetsage2 · abogen · vocalinux · ilya-timonin · musicgen · ai-music-copyright