Scoring contrasting scenes: 12 BPM apart and a new lead instrument

When scenes in one video should sound different, the Sume Music docs give three levers: change the broad genre family, set the tempo at least 12 BPM apart, and change the lead instrument. When the user wants one consistent score, do the opposite and keep all three. That is the whole rule, and it is cheap to test because each accepted Music generation costs a fixed $0.125.

An example set

The tempos below follow the 12 BPM rule: each neighbor differs by at least 12. The axes come from the docs' seven-axis brief; the values are examples, not recommendations.

Checking the 12 BPM gap

Neighboring scenes should be at least 12 BPM apart. 72 to 142 is a gap of 70, and 142 to 96 is 46, so both pass. If you moved the third scene to 84 BPM, the gap to the first would be 12 and still pass, but the gap to the second would be 58. Compare each neighbor, not only the first and last.

One consistent score instead

For one score across scenes, keep the genre, tempo and lead instrument, and change only the arc moment ("breakdown at 0:20"). If a scene still exists, pass it as image_url. End each prompt with "Instrumental, no vocals." and add "no spoken word" only under narration. Do not use negative_prompt: a non-empty value returns HTTP 400 with negative_prompt_unsupported.

• Tempo and key are written as numbers and names, not feelings.
• Instruments should carry texture: "Rhodes through tape wow", not "keys".
• The axes are directions, not guarantees. Listen to each artifact.

Budget

Three scenes at one take each is 3 x $0.125 = $0.375. Allow two takes for any scene that misses and the bill is still small. Use an Idempotency-Key per scene brief so a repeated request returns the same job.

After the first take

After a policy rejection, change the flagged content but keep the musical brief, and retry only inside the authorized budget. Do not reduce the request to a generic bed. If a take is close but wrong on one axis, change that axis only, so you learn which lever moved the result.