Settings

Performance

The AI models that MelonSound uses under the hood must be loaded into your computers memory when used. These models are very large, and require substantial space to fit.

This can sometimes lead to out of memory errors.

The main purpose of the performance settings is to allow users to control the memory pressure on their computer. To understand what is happening, think of it as a suitcase that you are packing things into:

Your Total Memory: 16 GB

- macOS + system services:     ~3.0 GB  (conservative)
─ MelonSound overhead:         ~1.0 GB
                               ────────
Available for models:          ~12.0 GB  (optimistic)
Safe budget:                   ~10.0 GB  (realistic)

When MelonSound starts up it will try to load the following:

VAE + Text Encoder:            ~1.0 GB (basic, always needed)
LM (Music Composer) Model:     ~2.4 GB (writes the music, "standard" quality)
DiT (Music Generator) Model:   ~4.5 GB (performs the music)
                               ────────
Total Space:                   ~7.9 GB

On top of this comes additional memory needed to actually compute things. Some settings can affect how much additional memory is needed. For example, longer songs will generally require more.

LM Quality

The LM model is the song writer. There are multiple models that can be used for this and quality can improve with heavier models.

If your computer has more than 16 GB of memory then you can probably set this setting to high which will tell MelonSound to load a higher quality LM model. This is why 24 GB of memory is recommended for MelonSound.

Off                         ~0.0 GB (completely disables the composer)
Standard                    ~2.4 GB (Standard quality)
High                        ~6.8 GB (Higher quality)

MelonSound can technically operate with no LM model at all, but quality may significantly decline.

GPU Memory Allocation

WARNING: This is a dangerous power setting.

The GPU memory allocation setting is a power user setting designed to let you disable certain safeguards for how much memory MelonSound's underlying engine is allowed to allocate on your system. This can give MelonSound additional headroom to grab extra memory from the system and allow it to complete operations in tight memory space. The downside is that the application can literally grab so much memory that your entire computer crashes. Use at your own risk.

Standard:                   Safe limits.
Elevated:                   2.2x system recommendation.
Unlimited:                  No limit.

Extras

These are extra features that can be enabled in MelonSound

Lyrics Generator

The built-in LM used for music composition is not capable of generating lyrics based on a literal prompt. The lyrics generator allows you to load a secondary LM that is configured to take specific prompts (e.g. A song about a cat who lives in New York City and loves tuna.) and turn it into lyrics.

The lyrics generator model requires additional memory of ~3.2 GB.