AI background music for a short video: a text brief or the clip itself
Two ways to get an instrumental bed for a short video: describe it in words, or let a model watch the clip. How to brief it and how to check the cut.

A short video with no music feels unfinished, and a stock track rarely fits: it is too long, it swells in the wrong place, and three other channels already use it. The studio has two ways to make a bed of your own. You can describe the music in a text brief and set the length yourself, or you can upload the clip and let a video-to-music model watch it and score it to the picture. Both are on the models page under Music. This post is about which one to reach for, how to brief it, and how to check the result before you publish.
One thing first, so the expectations are right: this is about an instrumental bed under picture. For sound effects that follow the action (steps, doors, wind), read the guide to sound for a silent clip. For a song with words, the music models that sing are a different job with a different brief.
Route 1: a text brief
Text-to-music models take a written description and, on most of them, a duration you set before you run. You never show them the video, so you are the one who knows where the cut is. That sounds like a drawback, and for a talking-head clip or a product loop it is an advantage: the picture has no story beats to follow, and what you need is a steady mood at a steady level that sits under a voice.
A brief that works names four things, in this order: the use (background under speech, an intro sting, a loop), the mood, the tempo and instruments, and the shape (where it builds, how it ends). Say instrumental out loud; several music models will add a voice if the prompt leaves room for one, and some have a switch for instrumental only under the advanced settings.
Instrumental background music for a calm product video with a voiceover. Warm and optimistic, around 95 BPM. Soft electric piano, muted bass, light brushed drums. No vocals, no big drops. Even level throughout, gentle fade at the end.Instrumental, tense and cinematic, slow build. Low pulsing synth bass and a ticking hi-hat at 120 BPM, strings entering halfway, rising to a hit on the final second, then silence.Set the duration to the length of your clip, or a second or two longer so you have room to trim. If your editor can loop audio, a short loop brief ('a seamless loop, no intro, no ending') is often cheaper than a long track, and it suits a looping header video well.
Route 2: the clip itself
A video-to-music model takes your video as the input and writes music that follows its mood, pacing and scene changes. The text prompt is optional and only steers the style. This is the route for footage with beats of its own: a travel montage with hard cuts, a reveal, a clip that slows down and then speeds up. A text brief cannot see a cut at the eighth second; this model can.
Light acoustic guitar and soft percussion, hopeful, no vocals.Give it the edit you will actually publish. If you trim two seconds afterwards, the music that was written to the old cut lands on the wrong frames. Lock the picture first, then score it.
When each route wins
- Voiceover, tutorial, interview, product loop: text brief. You want an even bed, and you control the length exactly.
- Montage with cuts, a reveal, a trailer: the clip. The model hears the edit and moves with it.
- A series of videos that must sound alike: text brief. Reuse one prompt and change only the duration, and every episode gets the same palette.
- You do not know what the clip wants: run the clip once with no prompt and listen. It is a quick way to hear a direction you can then describe in words.
Checking the cut
Lay the track under the picture in your editor and watch it three times, each time listening for one thing.
- The start and the end. Does the music begin on the first frame, and does it end, rather than stop? A hard stop in the middle of a phrase is the most common giveaway. A short fade in your editor fixes most of them.
- The voice. If anyone speaks, lower the music under the speech and listen on phone speakers, not only headphones. A bed that sounds right in a studio often buries a voice on a phone.
- The beats. On a scored clip, check that the big moments land on the cuts. If one is a few frames off, nudge the track; if the whole thing drifts, the picture probably changed after the run.
What the studio does not do
The music models return an audio file. The studio does not mix it into your video, duck it under a voice or fade it for you; that is a minute of work in any editor, and it keeps the levels in your hands. A text brief cannot see your cuts, and a scored clip only follows the picture it was given. Neither route promises a melody that matches a famous track, and a prompt that asks for one by name is the wrong prompt. If you are still making the video itself, the guide to picking a video model comes first; when the picture is locked, open the studio and score it.
Questions
Should I describe the music or upload my video?
Describe it for an even bed under speech or a series that must sound alike. Upload the video when the music should follow cuts, a reveal or changes in pace.
Will the music have vocals?
Only if you leave room for them. Write instrumental and no vocals in the brief, and use the instrumental switch where a model has one.
Do I get a video back with the music in it?
Not from these two routes. The music models return an audio file, and you place it under your clip in your own editor, where you also set the level and the fades. A separate video-to-video sound model returns the clip with new music and effects already in it, but then the mix is the model's, not yours.


