You type a prompt, and an audio file comes out. That is the picture most of us have of AI music. The YuE2 Studio that @superfang119 shows off does something different: the score comes out first. The creator writes (in Chinese) that it may be one of the very full-featured local AI music studios around right now 1, and says the most interesting part is that it does not simply spit out an MP3 2.
🤯 这可能是目前功能非常完整的本地 AI 音乐工作室之一!
— 方知微 (@superfang119) October 31, 2026
YuE2 Studio:把 AI 作曲、演唱、乐谱编辑和音频后期全部装进一个桌面软件。
最有意思的是,它不是简单地输入提示词,然后吐出一段 MP3。
🎼 先作曲,再演唱
输入音乐风格和歌词,AI 先生成旋律与和弦乐谱,再根据乐谱演唱成完整歌曲。
🎤…
Here is the flow. You describe a style and write the lyrics, the model composes a melody with chords as sheet music, and then it performs that score as a full song with vocals 3. According to @superfang119, you can edit notes, rhythm and key directly, keep the original compositional structure, and regenerate a different sung version 4. You like the melody but want a different delivery? You don't have to re-explain everything in words from scratch.
In the earlier piece, YuE2 Turns Your Desk into a Songwriting Studio, we read through examples made with YuE2. This time we look at the "studios" built around it.
YuE2 Studio: a studio in a single app
According to its README, YuE2 Studio is built on yue2.cpp, the native C++/CUDA port of YuE2, and ships as a standalone executable that needs neither Python nor Node.js 5. @superfang119 points to the same thing: a single desktop program starts it, with no Python or Node.js environment to set up yourself 6.
The feature list is long.
- Covers: it listens to a recording, writes its melody as a score, and has the model sing that melody in another style [[F587]]
- Stem separation into 6 stems: drums, bass, vocals, guitar, piano and the rest [[F588]]
- A built-in MCP server, so agents such as Claude Code can operate the studio [[F586]]
@superfang119 lists that last point too, as letting an AI agent control the music studio for you 7. You can picture it: you play the melody at your desk while an agent handles the export and re-sing steps.
On speed, the README has the author's own measurements. On their machine, with an RTX 4090 and the Q8_0 model set, a 1-minute song took 20 seconds and a 4-minute song took 54 seconds 8. These are numbers from the author's setup.
What you need is Windows 10/11 x64 9 and a GPU with 6 GB of VRAM or more 10. A Linux build exists, but the README calls it experimental 11.
One line worth knowing before you start: the YuE2 model weights are under CC BY-NC 4.0 with an additional creator permission 12. Individual creators can monetize their output for free, while companies using it commercially need to discuss a commercial license for the weights 13.
Game music, starting from a style or a hum
@songsong recommends YuE to game developers. The post opens by saying that if you wonder how to make music for your own game, this open-source project should help 14. @songsong writes (in Chinese) that YuE is a locally runnable AI music generation project: give it lyrics and a style and it produces a full song with vocals and accompaniment, and it supports instrumental-only tracks too 15.
如果你做游戏的时候想怎么给自己游戏来制作音乐
— 宋宋 (@songsong) November 9, 2026
那么这个开源项目应该可以帮到你
YuE 是一个可以在本地运行的 AI 音乐生成项目。 输入歌词和曲风,它能生成带人声、伴奏的完整歌曲,也支持纯器乐。
我们做游戏可以先找一个你希望的风格或者主旋律是自己哼出来的
然后再按乐谱生成音频… https://t.co/SZIJDT15Ie pic.twitter.com/hZsZPL9Mdk
The workflow @songsong describes is concrete. For a game, first find the style you want, or hum the main melody yourself, then generate the audio from the score 16. This is a recommendation, not an example @songsong made. Still, using a hummed melody as the starting point for a game soundtrack is easy to picture.
YuE2 also has native ComfyUI nodes 17. Our second studio lives right inside ComfyUI.
Plenio: a songwriting room inside ComfyUI
According to Plenio's README, the Plenio Music Production System is a song production suite for ComfyUI that goes from one sentence to a mastered song, using YuE2 or MiniMax Music 3 18. Version 1.0.0 came out on October 16 19. What follows is taken from the README for the current version.
- Everything runs on your own machine inside ComfyUI's own nodes, with no cloud service, no account and no API key [[F591]]
- The code is Apache-2.0 [[F590]]
- Lyrics are written by a local language model, or you can have it sing your own lyrics as they are [[F594]]
- Before the song is sung, a score editor lets you fix notes and chords [[F595]]
That last point, fixing it before it sings, is one of the highlights here. Listen to the melody that comes out, change the note that bothers you on the spot, then let it sing. The same idea as YuE2 Studio, putting the score at the center, reaches into ComfyUI's nodes as well.
The README also gives a sense of speed. On the author's 16 GB RTX 5060 Ti, a 1:44 instrumental, from writing to export, took 2½ minutes 20. You need ComfyUI 0.41.0 or newer 21.
Instead of rewording the prompt, fix the notes
What both studios share is a score in the middle. When something isn't right, you don't hunt for new prompt wording; you move a note, and the song gets sung again without losing its composition.
That makes you want tools for your hands. If you would rather play the melody in than click it with a mouse, a small MIDI keyboard helps.
Ad: the “Search on Amazon” buttons are affiliate links, and we may earn a commission. They open a search, not a specific listing.
And for hearing the breaths and smudges in a re-sung version, studio headphones are what you lean on.
If an RTX Spark desk is your plan
In the examples above, the authors measured speed on an RTX 4090 8 and an RTX 5060 Ti 20. Read those as facts about the creators' own gear. The Windows version the README lists is x64 9, and the README notes that the Linux build is experimental 11.
If you are weighing an RTX Spark machine for songwriting, check the official supported environments of the tools you want to run first.
Read next
- The earlier piece: YuE2 Turns Your Desk into a Songwriting Studio
- From song to video: Making a Music Video Solo
