Most slowdowns are temporary and easy to work around. Generation involves real compute, so demand and song length both matter.
Quick fixes
Retry after a moment if there is a queue, generate a shorter track or section, confirm your internet connection is stable, and check you have enough credits. If a generation appears stuck, refresh and try again.
- Retry during peak times.
- Try a shorter generation.
- Check connection and credits.
What is actually happening while you wait
A generation is not one step. The model first works out the musical metadata — tempo, key, structure, how long the song needs to be — then generates the audio itself, then masters and encodes the result. Most of the wall-clock time is the middle stage, and it scales with how long the song is.
The other variable is whether a GPU was already warm. The first request after a quiet period has to load the model weights before any audio can be produced, which adds a fixed overhead that later requests in the same session do not pay.
- Longer songs take proportionally longer — duration is the main driver.
- The first generation after an idle period pays a model-loading cost.
- Higher-fidelity models take longer than the faster ones.