1. Getting started
VuuWave turns a written brief into a finished audio track. You describe the style, supply lyrics, set length and key, and the local MusicCore engine generates the song on your own GPU or CPU. Nothing is uploaded for generation.
- Pick a workspace folder in the top-left. Finished songs are saved directly into it.
- Fill in Style / Prompt and Lyrics in the Studio panel.
- Select Submit. The first run loads the models, so it takes longer than later runs.
- Play the result from the Results list on the left.
2. Top bar
| Control | What it does |
|---|---|
| Workspace path | Select the folder for this project. Generated songs are written to the folder root as Song title.mp3; a duplicate name gets (2), (3), and so on. |
| Engine control (bolt) | Loads or unloads the MusicCore engine by hand. Red means models are missing. An outlined bolt means ready but unloaded — select it to load. A pulsing filled bolt means loaded — select it to unload and free the memory. |
| Settings (gear) | Opens Host Settings: language, appearance, transparency, background, MCP server, and MusicCore inference backend. |
3. Studio: describing the song
| Field | How to use it |
|---|---|
| Title | Becomes the output file name, so keep it filename-safe. |
| BPM | Tempo in beats per minute. Leave empty to let the model choose. |
| Key | For example C major or A minor. |
| Duration | Song length in seconds. Generation time scales roughly linearly with this value. |
| Style / Prompt | Genre, instrumentation, mood, production feel. Concrete beats vague: warm lo-fi hip hop, dusty piano, vinyl noise, relaxed works better than nice music. |
| Lyrics | The words that actually get sung. Write them out in full. |
| Pure instrumental | Turn on for a track with no vocals. |
| Reference audio | Optional. Guides the result toward an existing track's character. |
[write chorus here] is sung as those words. Leave the field empty, or enable Pure instrumental, if you do not want vocals.4. Results
The left panel lists everything generated in the current workspace. Each result offers:
| Action | What it does |
|---|---|
| Play / Pause / Stop | Preview the track in place, with a waveform drawn from the real audio. |
| Use as reference | Feeds this track back in as reference audio for the next generation. |
| Reuse prompt and parameters | Reloads the exact settings that produced this track into the Studio panel, ready to adjust. |
| Delete | Removes the result. |
| Refresh | Rescans the workspace folder. |
Generation settings are embedded inside each MP3 or WAV file, which is what makes Reuse prompt and parameters work. Re-exporting a file through a third-party audio editor can strip that data; the file still plays, but its settings can no longer be recovered.
Switching workspaces changes which folder the list reads. Songs in the previous workspace are left untouched.
5. Agent panel
The right panel is the shared VuMos Agent V17, the same assistant used across Vuu products. It can read the current draft, write lyrics, propose a style prompt, and fill in and submit a generation for you.
Useful requests
- “Write lyrics for a two-verse indie folk song about leaving a hometown.”
- “Turn my draft into a more specific style prompt.”
- “Fill in the studio fields for a 90 second instrumental and submit it.”
- “How do I stop generation from getting slow?”
The Agent can answer questions about VuuWave itself; it reads this guide when asked how something works. Its control bar offers model and mode selection, history, new chat, clear chat, and Agent Settings, which are shared with the same Agent in other Vuu products.
6. Inference backend and video memory
This section explains the one performance problem you are most likely to hit, and how to avoid and recover from it.
6.1 The symptom
Generation that normally finishes in seconds suddenly takes minutes. The GPU reports 100% usage, but the card stays cool and quiet and its power draw is low. Once this starts, it does not get better on its own — and it stays slow even after you unload the Agent model.
6.2 What is happening
Your graphics card has a fixed amount of fast video memory (VRAM). The music engine and the Agent language model both need a share of it. When their combined demand exceeds what the card has, Windows silently moves part of the data out of VRAM into ordinary system memory, which the GPU can only reach across a much slower connection. The work still runs on the GPU — it just spends most of its time waiting for data. That is the low-power, high-usage pattern.
6.3 How to avoid it
- Mind the total, not each part. A 12 GB card comfortably runs the music engine alongside a 1B Agent model. A 2B model pushes it to the edge; larger models will exceed it. Cards with 16 GB or more have considerably more headroom.
- Load the music engine first. Whichever component claims video memory first gets the fast memory. Selecting the bolt icon before loading a large Agent model gives the engine the better position.
- Unload what you are not using. The bolt icon frees the engine's memory immediately. There is no automatic timeout.
- Or remove the uncertainty: switch to the CPU backend. See 6.5.
VuuWave will refuse to start a generation when free video memory is critically low, and tells you so rather than starting a run that would crawl. This is a guard, not a guarantee — a job that fits at load time can still exceed the card at its peak.
6.4 How to recover once it is slow
- Open Settings → MUSICCORE INFERENCE and select Restart kernel. This is the quickest route.
- If that does not help, close VuuWave completely and reopen it.
- If generation is still slow after a full restart, the graphics driver itself may be in a bad state. Press Win+Ctrl+Shift+B to restart the display driver — the screen blinks once, open applications are unaffected.
6.5 Choosing a backend
Under Settings → MUSICCORE INFERENCE you choose how the music engine runs. The choice is fixed while the engine runs, so changing it restarts the kernel.
| Backend | Character | Choose it when |
|---|---|---|
| Vulkan (GPU) | Much faster, but depends on having enough video memory. If it runs short, speed collapses as described above. | Default. Your card has room for the models you use. |
| CPU | Slower, but completely predictable — the same request always takes about the same time, scaling with the song's length. Uses no video memory at all. | Video memory is tight, you are running a large Agent model, or you want a run that cannot degrade partway through. |
Measured on one test machine (RTX 5070 12 GB, 24-core CPU) as a rough sense of scale, not a specification: a 30 second song took about 6 seconds on Vulkan and about 53 seconds on CPU; a 200 second song took about 7 minutes on CPU. Your results will differ with hardware.
6.6 The two Vulkan options
| Option | What it means |
|---|---|
| Bias the driver against demoting our VRAM | Leave on. It asks the graphics driver to move some other application's data out first, if something has to be moved. It is only a request — it cannot reserve memory, take memory back from another program, or undo a slowdown that already happened. |
| Allow spilling into shared memory | Leave off. Off means an oversized model simply fails to load, with a clear error, instead of loading and then running many times slower. Turn it on only when you deliberately want to run a model too large for your card and accept the speed penalty. |
7. Settings
- Language: English, 中文, Español, Français. The whole window updates immediately.
- Appearance and transparency: theme and per-region transparency.
- Background: default wallpaper, a local image, a solid color, or none.
- MCP server: external tool endpoint shared with other Vuu hosts.
- MusicCore inference: backend selection, covered in section 6.
- Version: the running VuuWave version.
8. Common questions
Generation suddenly became very slow.
Almost always video memory. See section 6 — restart the kernel from Settings to recover, then reduce what is loaded at once or switch to the CPU backend.
The bolt icon is red.
The model files are not installed yet. VuuWave opens the MusicCore resource download page automatically; download the separate GenPack and wait for the progress bar and extraction to finish.
Why does the first generation take longer?
Models load on first use. Later runs in the same session reuse them.
Closing VuuWave left memory occupied.
Close the window normally rather than ending the process from Task Manager. A normal close unloads the engine and shuts the kernel down; force-ending it mid-generation can leave the graphics driver holding memory, which then needs a driver or system restart.
Can I run VuuWave on a server without a GPU?
Not practically. The models need roughly 8 GB of memory just to load, and the CPU backend on a fast desktop processor takes minutes per song — a small virtual server cannot do it at all.
My song file lost its settings.
Re-exporting through a third-party audio editor discards the embedded generation data. Keep the original file if you want to reuse its settings.
How do I get help?
Ask in the Agent panel on the right. It can read this guide and answer questions about how VuuWave works.