Write music video prompts that hold their visual idea: shot lists, camera language, lighting, consistency fixes, beat-synced edits, and pitfalls.
A music video prompt is not a summary of a song. It is a dense, repeatable instruction that tells an AI video generator what the camera sees, how it moves, what the light is doing, and how long the moment should last. Weak prompts fail for a predictable reason: they describe a mood and stop. "Dreamy, emotional, cinematic" gives a model almost nothing to build, so it averages your idea into a generic look. "A single overhead bulb swings across a wet parking garage while a singer in a silver bomber jacket walks toward a handheld camera, 35mm, shallow focus, sodium light" gives it a scene.
The difference is craft, not vocabulary. What follows is a full working method: reading the track for structure, choosing a concept that survives a dozen separate generations, writing shot-level prompts that hold together, and fixing the problems that show up in the first rough cut.
Start With the Song, Not the Prompt
Before you open a generator, listen to the track three times with something to write on. The first pass is for feeling. The second pass is for structure. The third pass is for images the lyrics already contain.
On the structure pass, write down timestamps for the intro, each verse, each pre-chorus lift, each chorus, the bridge, and the outro. Note where the drums drop out, where the vocal doubles, where a synth pad enters. These timestamps become your shot boundaries. A music video that cuts on structural changes feels intentional even when the images are simple; a video that cuts on arbitrary intervals feels random no matter how beautiful the frames are.
A quick example map for a mid-tempo track:
- 0:00–0:12 intro: one wide shot, no movement, ambient sound
- 0:12–0:41 verse: three medium shots, slow drift, cool grade
- 0:41–0:55 pre-chorus: handheld, tighter framing, warmer light entering
- 0:55–1:24 chorus: wide + close alternation, faster camera moves, saturated color
The third pass is where most of your prompt material comes from. Circle any concrete noun in the lyrics — a name, a place, an object, a weather condition. Concrete nouns are anchors you can describe visually. Abstract lines are better translated into a physical action than illustrated literally.
Finally, choose one visual idea and protect it. A music video concept needs to survive ten to fifteen separate generations, each of which is a small independent decision. If your idea needs twelve specific shots that only work in one exact order, you will not get it. If your idea is "a figure keeps returning to a room that rearranges itself while they are gone," you can express it in almost any shot you end up with.
What a Music Video Prompt Actually Needs to Do
Text-to-video models weight nouns, verbs, and camera terms far more heavily than adjectives. Adjectives get averaged into a house style. If you write "moody," you get the model's default moody. If you write "lit only by a flickering vending machine," you get something specific, because you have described a physical light source and its behavior.
The four load-bearing parts
Almost every strong shot prompt answers four questions in order:
- Subject and wardrobe. Who or what is on screen, and what exactly are they wearing or holding? Wardrobe is the fastest consistency lever you have.
- Action with a beginning and an end. Not "dancing" but "steps forward, turns, and drops out of frame." A shot needs a before and an after, otherwise the model invents a drift.
- Camera. Framing, height, lens feel, and movement. "Low angle, 24mm, slow dolly left" is a decision. "Cinematic camera" is not.
- Light and texture. Where the light comes from, how hard it is, what is in the air, and what the grade looks like.
Compare these two prompts for the same moment:
Weak: A cool music video shot of a singer in a city at night, cinematic, emotional, 4k.
Strong: Medium close-up of a singer in a black leather coat, collar up, walking toward camera through a narrow alley, 40mm, slow handheld push-in that keeps her centered, lit by a single pale streetlamp above and behind her plus a soft bounce from wet asphalt, thin drizzle, cool blue-teal grade, gentle halation on highlights.
The strong version is longer, but every clause is doing work. Nothing can be guessed wrong without changing the shot.
Words that do real work
Build a personal list of terms that reliably change output: lens millimetres, camera height, movement verbs (push in, pull back, orbit, crane, whip pan, static lock-off), light sources (sodium vapor, fluorescent tube, bare bulb, firelight, overcast daylight), atmosphere (haze, dust, steam, rain, smoke), and texture (16mm grain, anamorphic flare, soft-focus diffusion). Replace every abstract adjective with one of these.
Build a Concept That Survives a Dozen Shots
A useful test: can you describe your video in one sentence that does not mention the song? If yes, you have a concept. If you need three sentences, you have a shot list, which is a different thing.
Strong one-line concepts for generated music videos tend to share three traits: a repeatable location, a repeatable subject, and one visible rule that changes over time.
- "A dancer performs the same routine in four rooms that get progressively emptier."
- "A courier walks the same corridor, and each pass the walls are a different material."
- "A band plays inside a slowly rotating room, gravity staying normal."
The visible rule is what gives the edit meaning. Without it, you have a collection of pretty clips; with it, the audience starts anticipating the next change.
Define a motif kit before you generate anything
Write down a small set of elements you will reuse in every prompt: two wardrobe pieces, one prop, one light source, one color accent, one texture. For example: silver bomber jacket, black boots, a payphone handset, a green fluorescent tube, teal shadows, 16mm grain.
Because you reuse the exact same words for these elements in every prompt, the model receives a consistent signal. This single habit fixes more continuity problems than any post-production trick.
The Shot-Level Prompt Template
Here is a template you can fill in for each timestamp on your structure map. Keep the clause order stable so you can compare results and find which part is failing.
[shot size] of [subject + wardrobe + prop], [action with a clear start and end],
in [location with two concrete details], [camera height + lens feel + movement],
lit by [light source + quality + direction], [atmosphere], [grade and texture],
[duration or pace note].
A worked example
Say your chorus lands at 0:55 and you want a lift. Fill the template:
Wide shot of a singer in a silver bomber jacket holding a payphone handset, stepping out of a doorway and turning to face the street, in an empty laundromat parking lot with cracked asphalt and a flickering green sign, camera at chest height, 28mm, slow crane up as she turns, lit by the green sign and a single headlight sweeping across frame, light haze, cool grade with warm skin tones, one continuous take that ends as she lifts the handset.
Now the verse version, same subject and location, lower energy:
Medium shot of a singer in a silver bomber jacket standing still beside a payphone, shoulders dropping as she exhales, in an empty laundromat parking lot with cracked asphalt and a flickering green sign, camera at eye height, 50mm, locked off with a barely perceptible drift, lit by the green sign only, light haze, cool grade, slow and quiet pace.
Same words, same motif kit, different energy. That is how a generated video starts to feel like one video instead of twelve.
Translate tempo into duration
Fast songs do not need fast camera moves in every shot. They need shorter clips. A practical rule: calculate roughly one beat in seconds (60 divided by BPM). At 120 BPM a beat is half a second, so a two-bar phrase is four seconds. Generate clips slightly longer than the phrase you need — six seconds for a four-second slot — and trim in the edit. That gives you handles for matching motion at the cut.
Camera Language That Changes the Feeling
The camera is the cheapest emotional control you have. A locked-off wide shot reads as observation; a handheld close-up reads as intimacy; an orbiting camera reads as unease or euphoria depending on speed.
| Movement | What it signals | Good for |
|---|---|---|
| Static lock-off | Composure, formality | Intros, final chorus reveals |
| Slow push-in | Growing attention | Pre-chorus, emotional lines |
| Handheld follow | Immediacy, chaos | Choruses, performance shots |
| Crane up | Release, scale | Drops and key transitions |
| Whip pan | Restlessness | Bridge sections, montage |
| Orbit | Obsession, disorientation | When the lyric repeats |
Two practical notes. First, one movement per shot. Models frequently blend two movements into a wobble when you ask for both. Second, name the end state of the movement if it matters — "push-in that settles on her hands" gives the shot somewhere to land.
Light, Texture, and Color Grade
Lighting is where generated footage most often reveals itself. Flat, even, source-less light looks synthetic. Give every shot a visible origin: a window, a lamp, a screen, a fire, a headlight, a phone.
State the quality and the direction together. "Soft window light from frame left" is actionable. "Nice lighting" is not. Add one imperfection — haze, dust, drizzle, lens flare, a slightly overexposed highlight — because imperfection reads as photographic.
For grade, pick two anchors and a texture: teal shadows with warm skin, or bleached highlights with crushed blacks, plus grain or halation. Repeating the same grade sentence across all prompts is what makes separately generated clips cut together without heavy correction.
Keeping Faces and Places Consistent
Full consistency across many generations is still the hardest part of AI filmmaking. Plan around it rather than fighting it.
- Create reference stills first. Generate a character portrait and a location plate with an AI image generator, then use those images as the visual anchor for video generations.
- Freeze your descriptor strings. Copy and paste the exact wardrobe and location sentence into every prompt. Do not paraphrase between shots; small changes compound.
- Vary shot size, not subject. A wide, a medium, and a close-up of the same person read as coverage. Three medium shots of slightly different people read as a mistake.
- Hide faces when you cannot hold them. Over-the-shoulder, silhouettes, hands, reflections, back-of-head shots. Music videos tolerate this far better than narrative films do.
- Reuse locations instead of inventing new ones. Four locations is generous. Two is comfortable.
Syncing Cuts to the Beat: A Practical Workflow
Once you have clips, the edit determines whether the video feels musical. A repeatable sequence:
- Drop the track on the timeline and mark the structure. Add markers at every section change and every drum accent you care about.
- Place your strongest shot on the first chorus downbeat. Build the rest of the edit around that anchor.
- Cut on motion, not just on the beat. If a subject is walking left at the cut, the next shot should continue leftward movement, or contrast it deliberately.
- Alternate scale. Wide, close, wide, medium. Two adjacent shots of the same scale feel like a stutter.
- Hold longer than feels comfortable in verses, cut faster in choruses. The contrast is the whole effect.
- Add one hard cut on a silent beat. A single frame of black or a jarring cut right before a drop costs nothing and reads as intent.
- Grade the whole timeline last, with one look applied across all clips, then trim individual shots that refuse to match.
For a tighter start, browse video templates that already follow a musical structure, then replace the prompts with your own motif kit.
Mistakes That Wreck Otherwise Good Prompts
- Describing a mood instead of a moment. Fix: convert each adjective into a light source, a lens, a movement, or a texture.
- Asking for a story in one clip. A single generation cannot carry a beginning, a turn, and a resolution. Fix: one action per shot, assembled in the edit.
- Changing wording between shots. Even small paraphrases shift wardrobe and lighting. Fix: copy exact strings.
- Ignoring the aspect ratio and delivery format. Vertical crops differently. Fix: write prompts for your final frame, and check framing at that ratio.
- Overloading with contradictory light. "Golden hour" plus "neon night" produces mush. Fix: one dominant source per shot.
- Generating exactly the length you need. No handles means no room to match motion. Fix: generate long, trim short.
- Using the same shot size everywhere. Fix: write the shot size as the first word of every prompt.
- Skipping the concept test. If you cannot state the idea in one sentence, the edit will not fix that. Fix: write the sentence before the shot list.
Building a Reusable Prompt Library
After two or three videos you will notice you keep rewriting the same clauses. Save them. Build a plain text file with four blocks: motif (wardrobe, props, location details), camera (movements you like at specific energies), light (sources and qualities that worked), and grade (two anchors plus texture). Then a new prompt is assembly, not invention.
Keep a version note next to each block — which generator, which aspect ratio, what it produced. Model behavior shifts between versions, and a block that worked last month may need a small correction. Public prompt libraries are useful for vocabulary, but your own tested blocks will always outperform borrowed ones because they match your motif kit. More workflow breakdowns live on the Orelon blog.
FAQ
How long should each generated clip be?
Generate six to ten seconds for a slot that will end up three to five seconds in the edit. The extra length gives you handles to match motion at both ends of the cut. Avoid generating exactly the final duration, since you will have no room to trim a bad first frame.
Do I need a different prompt for every shot?
Yes, but most of the text should be identical. Keep wardrobe, location details, light source, and grade word-for-word the same, and change only shot size, action, and camera movement. Consistency comes from repetition, variety comes from three clauses.
Can I generate a full music video from one prompt?
You can generate an interesting clip, but not a structured video. Sections, drops, and transitions need separate generations placed on a timeline. Treat the model as a shot factory, not an editor.
How do I stop faces from changing between shots?
Start from a reference still, reuse the exact same character description in every prompt, cover with different shot sizes, and consider shots that hide the face. When a face must be clear and consistent, limit how many times it appears and place those shots at emotional peaks.
What about lyrics — should prompts include them?
Use lyrics as a source of concrete nouns, not as text to be rendered. Asking a model to display words on screen usually produces mangled lettering. Translate a lyric into an object, a gesture, or a location instead.
How many locations can I realistically hold together?
Two locations used well will look more intentional than five used loosely. If the concept needs a third, introduce it at a structural moment — a bridge or a final chorus — so the change reads as a decision rather than an error.
Turn Your Shot List Into Footage
A music video prompt is a contract between the song and the frame: subject, action, camera, light, texture. Write the one-line concept first, build a motif kit, keep your descriptor strings identical, and generate longer than you need. Everything else — the grade, the cut rhythm, the moment the chorus lands — is assembly work you can control.
When you are ready to move from notes to frames, the Orelon AI video generator is built for exactly this loop: write a shot-level prompt, generate, compare, keep the block that worked. Start with one chorus, get it right, then build the rest of the timeline around it.

