Sound is where audiences forgive nothing. A slightly-off AI image reads as style; a slightly-off AI score reads as cheap. We learned this scoring a 29-second action tease — helicopter, dragons, golden hour — and it took three generations to get something shippable. Here’s the honest ledger.

Attempt 1: “epic orchestral trailer music”

We asked for driving percussion, staccato strings, brass swells, a heroic theme. What we got was technically all of those things — and it sounded like a family holiday film where they’re racing to deliver presents before the kids wake up. Bright, bouncy, safe.

Lesson: text-to-music models default to pleasant. “Epic” in the training data is apparently mostly uplifting epic. If you want menace, you have to ask for menace explicitly.

Attempt 2: darker, in a minor key

Braams, war drums, low string ostinato, “no bells, no festive elements.” Better — genuinely darker, and this is the one we shipped. Still noticeably AI in its structure: it builds the same way every AI track builds, in tidy eight-bar blocks with no surprises.

Attempt 3: percussion-only sound design

Our theory: melody is where the corny lives, so remove the melody. Pure trailer sound design — sub drones, ticking, hits. The result contained piercing high-treble artifacts between the percussion hits that were genuinely painful on headphones. Straight into the bin.

Lesson: instrumental sparseness exposes the model. With melody and full arrangement, artifacts hide in the mix. Strip the track down and every synthesis flaw is naked.

What we’re doing about it

  1. Ship the best-of-three and move on. Music is one layer of a video; a 7/10 score under fast cuts and sound effects reads fine. Perfectionism here is a trap.
  2. Watch Suno. Their API partner program just opened applications, and their creator program pays short-form creators directly. The quality gap between dedicated music models and general-purpose ones is real — this is the upgrade path we’re pursuing.
  3. Budget for re-rolls. Each generation costs pennies. Generate three, pick with your ears, and never ship the first roll.

The bigger principle: sound matters to people more than they can articulate. Viewers will watch AI-generated dragons without blinking, then close the tab over a shrill hi-hat. Spend your quality attention where the audience’s tolerance is lowest.

The tease this scored ends our latest episode. The full AI-filmmaking breakdown is next.