A music prompt works like a brief to a session player: a few concrete words about genre, mood, instruments and sound get you further than a paragraph of feelings. The same brief works for most text-to-music models; what differs is where the tempo goes.
The advice here is not specific to one app. ACE-Step 1.5 is an open model, from the ACE-Step 1.5 repository. If you already run it locally, the same prompts apply there.
The five things a good prompt names
| Detail | Weak | Strong |
|---|---|---|
| Genre | "electronic" | "deep house" |
| Mood | "nice" | "late-night, melancholic" |
| Instruments | "music" | "warm Rhodes, rolling bassline, soft shaker" |
| Sound and production | "lo-fi, tape saturation, close and dry" | |
| Vocals (songs only) | "breathy female vocals, stacked harmonies in the chorus" |
Put them together and you have a prompt: Deep house, late-night and melancholic, warm Rhodes, rolling bassline, soft shaker, close and dry.
An era helps too, because it brings a whole sound with it: "80s synth-pop", "90s boom bap", "70s soul".
What to leave out
- Contradictions. "Calm, aggressive drum and bass" makes the model pick one, and not always the one you meant.
- Instructions about structure in an instrumental prompt. "Starts quiet then
builds" sometimes works; for songs, the structure belongs in the lyrics tags
(
[Build],[Drop]) or in the Studio's timeline. - Numbers for ACE-Step. Tempo, key and time signature go in their own fields. Written into the prompt as well, they compete.
Tempo: in its own field
Some text-to-music models read the tempo from the prompt. ACE-Step 1.5, which the tool runs, takes it from the Tempo (BPM) field under Advanced, along with the key and time signature. Leave them on Auto, or fill them in, and keep the prompt for the sound.
Instrumentals: switch the vocals off
For music without singing, open Lyrics and turn on Instrumental (no vocals). Leaving the lyrics empty does the same. Then leave vocals out of the prompt too, so it does not ask for something the switch has turned off.
- Start with genre, mood and two or three instruments. Generate a short track, 15 to 30 seconds, so each try is quick.
- Turn on Keep this seed under Advanced once you hear something close.
- Change one thing at a time: a word in the prompt, the tempo, an instrument. With the seed kept, the difference you hear is that change.
- When it is right, set the full length and generate the final version.
A few that work
These are starting points to edit, not recipes:
Warm lo-fi beat, mellow piano chords, vinyl crackle, relaxed
Cinematic strings building to big taiko drums, heroic, wide
Dreamy ambient pads, slow and wide, distant piano, no drums
Upbeat synth-pop, bright arpeggios, driving drums, female vocals
Gritty garage rock, fuzzy guitars, shouted chorus, live room sound
The first three are instrumentals: turn on Instrumental (no vocals). The last two are songs and need lyrics.
