A single prompt gives you a whole song or nothing. Arranging it gives you control in the two places that matter: before generating, where the parts go and how long they are; after generating, which parts to redo.
ACE-Step 1.5 supports repainting a region of a song and extending it by repainting past the end. Its own app, from the ACE-Step 1.5 repository, exposes these with every checkpoint if you run it with Python on a strong NVIDIA card. We have not run that install ourselves, so follow the repository's README.
Plan the song before you generate
Switch the tool to Studio. The song appears as blocks on a bar grid: intro, verse, chorus and so on. Start from a template, or from lyrics you already wrote in the Simple view.
- Drag a block to move it. A marker shows where it will land.
- Drag a block's right edge to make it longer or shorter, in whole bars.
- Drag a section type from the row above the timeline onto it to add one, or click it to add after the selected block.
- Select a block to give it lyrics, a short style ("tender", "anthemic") and an energy level.
- Add a moment (a build, a drop, harmonies, a fade) at a bar inside a section. It travels with its section when you move it.
The tempo turns bars into seconds. The length readout shows the whole plan against the model's limit of three minutes, and Generate waits until it fits.
The plan becomes tagged lyrics, in order, plus the tempo, key, time signature and length. That is all the model reads. It follows the order closely, and the tempo keeps the parts near their planned length, but it picks the exact bar.
Let the planner fill the gaps
Plan with AI runs ACE-Step's own song planner over your prompt and lyrics. It fills in a tempo, key and time signature, and may suggest a fuller prompt, which you can take or leave. It is a separate 1.4 GB download the first time, and Undo takes back what it changed.
Fix what came out wrong
After generating, the take sits under the plan on the same time axis, with the planned section names above the waveform. That makes the misses easy to see.
- Drag across the waveform to select the part, or click a section name above it to select that section. With Snap to beats on, the edges land on a beat.
- Optionally write a prompt just for this part: "add a gospel choir", "half-time drums". Empty, it uses the song's prompt and the section's style.
- Choose how much to change. Keep more stays close to what was there; Change more lets the model take it further.
- Press Regenerate selection. The result is a new take; the old one stays in the list.
- Use Compare with previous to switch between the two at the same position while it plays.
If the words of that part need to change, edit the section's lyrics first. The model reads the lyrics of the whole song when it fills in a part.
Make it longer
Extend adds 10, 20 or 30 seconds after the end, continuing the take. Or pick a section type: the plan gets that section at the end, and the song is extended by its length. Build a long song this way, a section at a time, and listen before each step. The total stays within three minutes.
