A soundtrack good enough to sell, and wrong for its scene

I generated a track for a chapter in the second phase of Perpetūra and it was good. Not “fine for a placeholder” good — proper structure, an arrangement that knew when to pull back and let a line breathe. If I’d commissioned it from a composer I’d have paid for it and been glad. I dropped it under the scene it was written for and it was wrong within four bars.

Not wrong in any way I could point to on its own. Wrong for that scene, at that moment. The track wanted to be the emotional center of the room; the scene needed to stay quiet and let the dialogue carry. Suno had written something that earned attention. The scene needed something that got out of the way.

How these actually get made

The soundtrack starts as a prompt to Suno: mood, tempo, instrumentation, sometimes a line of the actual prose fed in as the lyric so the song is singing the scene and not something adjacent to it. What comes back is a batch, four or five takes on the same brief, and I listen to all of them before I decide anything. Scene art goes through an image model the same way: framing, light, what the character is doing with their hands, and it comes back in batches too. The volume is the new part. I’m not commissioning one piece and hoping. I’m making twenty and keeping one.

The gap the tools can’t close

Scene art fails a different way than music. The first pass can be gorgeous, good light and real weight to the composition, and then you look again and the face has drifted half a degree from the reference sheet, or a hand is doing something the character wouldn’t, or the model has invented a background detail the scene never established. None of it shows up on the first look. It shows up once you already know what the scene is supposed to be.

That’s the part the tools can’t do for you: they don’t know the scene, they know the prompt. A composer I’d hire would ask what the beat needs before writing a note. Suno doesn’t ask. It answers the prompt well, and it’s on me to notice the answer is for a different question than the one the scene is asking.

What closes that gap isn’t a better prompt. It’s generating more than I need and rejecting most of it: twenty tracks to find the one that recedes instead of the one that impresses, a dozen images to find the one that matches what’s already established instead of the one that looks best alone. That only works if I know the scene well enough to say no to nineteen good options without flinching. Which is the old skill, closer to what an editor or a music supervisor does than a composer or an illustrator. The tool didn’t hand it to me. I had to bring it.

Production got cheap; judgment didn’t

I want to be honest about the other half of this, because the fit problem can read as a complaint and it isn’t one. Suno is startlingly good. A year ago the alternative to a commission was no music, or a stock track that fit nothing in particular. Now the floor is commission-quality in minutes, and I’d be lying to pretend that isn’t a real change in what one person can make alone.

“AI is going to change the creative industries” is true and says almost nothing. The version I act on is narrower: the cost of a first draft in music and image has gone to roughly zero, so the whole value of a solo project moves into the parts that were never about production — knowing what the scene needs, and having the taste to recognize the right take when it finally shows up at number fourteen. That’s the thing I keep telling myself going into the next round. Cheap generation is very good at talking you out of sitting with the scene long enough to know what it needs before you go looking.