Negative Prompts
Saying what you do not want often works better than repeating what you do.
Related
The problem
Models like to complete: ask for strings and drums may appear, ask for quiet and a pad may creep in. Repeating "not too busy" rarely helps, because negation is easily lost in language models. Negative prompts list exclusions separately and work more directly.
How to write them
Be concrete: "no electronic drum machine" beats "not too modern". List three to five items, since more start to conflict. Common exclusions include vocals, drum kit, distortion, heavy reverb and sudden modulations. Exclude categories rather than single notes; "not that high note" is not actionable.
Practice
Write the positive description first, then add exclusions. After generation, listen for anything excluded that survived and strengthen that line. Exclusions can overshoot: removing reverb dries the sound out and removing low end thins it. Prune exclusions you no longer need so the prompt does not grow without limit.
What to listen for
On the funk kit, whether the drum machine is truly gone; in the hymn, whether vocals have crept in; in the traditional tune, whether a modern kit has vanished. Good exclusions give the positive description room.
Listen in the library
All tracks from midicn-lib — inline badge is the license tier (C1 commercial / C2 non-commercial / C3 study only).
Sources
- 反向条件在生成模型中的实现，依公开模型说明核对，未抄原文
- 排除项与正面描述的关系，依创作实践通识核对，未抄文本
Used for fact-checking only; all prose is original.