How to Make an AI Music Video: Shot Lists, Lip Sync and Cutting on the Beat

ClipSpeedAI Oct 8, 2026 ◷ 14 min read Updated Oct 8, 2026

What do you need before you generate the first shot?

A music video is cut to fixed audio, so lock it first. Change the edit after you've paid for shots and you pay again.

• The final master, plus the isolated lead vocal if you can export it; lip-sync tools read the voice.
• The tempo and a section map with start and end times.
• Permission to use the music in a video.
• Artist references: a headshot and a full-body image, plus written consent if the performer is real.
• A release format: a 16:9 master, or vertical-first for Shorts, TikTok and Reels.

Beat math for shot lengths

One beat lasts 60 ÷ BPM seconds, and a 4/4 bar holds four. At 120 BPM that's 2 seconds per bar and 16 seconds for an eight-bar chorus; at 24 fps each beat is exactly 12 frames. Match bars to what each model returns per request: Veo 3.1 makes 4, 6 or 8 seconds, Kling 3.0 3 to 15 and Seedance 2.5 4 to 30. So one Veo clip covers four bars, and one Seedance request can cover a whole chorus. Generate slightly long and trim both ends to a beat.

How do you turn the song's structure into a shot list?

Give each section one job. Lengths assume 120 BPM.

Two habits keep the budget sane. Reuse setups: the chorus comes round two or three times, so generate one set of performance shots with spare takes and cut them differently each time. Spend on the chorus: it's what a cut-down opens on, so it earns the lip-synced close-ups; verses can lean on b-roll.

How do you keep the artist the same in every shot?

1. Build a reference pack: one headshot and one full-body image, same wardrobe, neutral light.
2. Write a lock list for every prompt: hair, wardrobe and one detail nobody could miss, such as a single gold earring.
3. Make a start frame for each scene from the same references, so each scene re-anchors the face instead of inheriting drift.
4. Use character tools where they exist; Kling's 3.0 guide describes binding a subject's look, and a voice tone, to keep it consistent.

Seedance 2.5 doesn't accept directly uploaded references containing real human faces, and Veo 3.1 allows only adults in image-to-video and reference-image generation. Neither replaces the artist's written permission. More in our guide to consistent AI characters.

Which lip-sync route fits your song?

The three routes differ on the question that matters most: is the voice you hear your actual recording?

Prices and limits read October 8, 2026 on each provider's own pricing page, model page or docs. USD per second, list rates; sync.so's ranges are at 25 fps and vary by plan.

Why native audio isn't lip sync to your track

Seedance 2.5 and Veo 3.1 generate sound with the picture. Put lyrics in the prompt and the character sings them, in a voice and melody the model invents; BytePlus's Seedance 2.5 tutorial includes a rap music video made this way. That suits a concept piece. It breaks when the release is your master: lay the real vocal over the clip and the mouth follows a different melody. Seedance 2.5 does take audio references (up to 10 clips, 30 seconds in total) for music, voice or timbre, but treat them as direction, not a track it will reproduce. More in our native audio guide.

The audio-driven route exists for singing: ByteDance's OmniHuman 1.5 page describes making a digital singer from a single image and a song, with separate tracks routed to each character in group shots.

What does 30 seconds of on-screen singing cost?

One attempt at list price, at 720p where the host offers a choice:

Our arithmetic from the list rates above: one attempt, before taxes, plan fees and retakes.

The real cost is rate times attempts, and lip sync raises attempts because a take can look right and still miss the words. Draft low, judge mouth shapes on the hard consonants, then render finals. See our AI video cost breakdown.

Copy-ready prompts for one chorus

"Last Stop Kingdom" is synth-pop at 120 BPM with a 16-second chorus. Swap in your own lyrics and references.

1. Full-body reference for the artist (any image model)

Full-body studio reference photo of an invented pop singer named Juno, standing straight and facing the camera, arms relaxed at her sides. Woman in her late twenties, tall and slim, warm brown skin, short platinum-blonde finger waves, one small gold hoop in her left ear only. Wardrobe: silver sequined jumpsuit with flared legs, white platform boots, a red coiled microphone cable looped over her right shoulder. Plain mid-grey seamless background, soft even light from the front. Neutral expression, mouth gently closed. Vertical 9:16, sharp focus head to toe, photoreal. No text, no logos, no other props.

Run it again as a head-and-shoulders portrait for the headshot. A closed, relaxed mouth gives lip sync a clean starting shape.

2. Chorus performance (Seedance 2.5)

Seedance reads {} as dialogue, () as music and <> as sound effects, so sung lines sit in braces.

Chorus performance for the song "Last Stop Kingdom". 16 seconds, 9:16, four shots with hard cuts on the bar lines at 120 BPM.
Image 1 is Juno's face; Image 2 is Juno's full body and outfit. Use face, hair, earring and wardrobe only; ignore the grey background and studio light. Juno is identical in every shot.
Setting: the open roof of a moving red double-decker bus at night, crossing a wide avenue lined with neon shop signs.
Shot 1 (0-4s): wide from the rear corner of the roof, static. Juno stands centre roof, mic in her right hand, and sings line one.
Shot 2 (4-8s): medium close-up from the front, slow push-in. She leans toward the lens and sings line two, eyes on camera.
Shot 3 (8-12s): low angle from roof level, static. She throws her left arm up on the downbeat as the bus passes under a railway bridge.
Shot 4 (12-16s): close-up, handheld with a slight sway. She sings line three and closes her eyes on the last word.
Wind from the front pushes her hair and the cable backward; the mic never leaves her hand.
Light: magenta and cyan neon from both sides of the street, warm streetlights passing overhead; her face stays lit.
Juno sings in English, lips matching every word: {We ride the night like it owes us a crown} {Last stop kingdom, nobody gets off now} {Hold on tight, hold on tight}. (Synth-pop, 120 BPM, punchy kick on every beat.) <bus brakes hiss> on the final bar. No other voices.
Photoreal music video, 35mm film grain, saturated colour. No subtitles, no on-screen text, no logos.

3. Verse b-roll with no vocals (Veo 3.1)

Eight-second b-roll for a music video verse. Inside the lower deck of a night bus, a rain-streaked window fills the left of frame; neon shop signs slide past outside in long magenta and cyan streaks. A commuter in a beige raincoat rests her head against the glass and taps two fingers on her knee in steady time. Camera locked off at seat height, 35mm look, shallow depth of field, focus on her hand. Cold fluorescent strip light overhead, coloured neon washing across her face from outside. Sound: engine hum, rain on glass, an indicator ticking. No music, no singing, no dialogue.

Mute it under the master; ambience only keeps a stray melody out of your ears while you cut. If a Seedance shot adds unwanted music, BytePlus suggests listing every synonym (music, BGM, score, instrumental, melody) at the start and end of the prompt.

4. Your real vocal on a still (OmniHuman 1.5 or Kling AI Avatar)

Image: Juno's headshot, facing the lens, mouth gently closed.
Audio: the isolated lead vocal for the chorus, 16 seconds, no instruments.
Prompt: Juno sings with full commitment, eyebrows lifting on the long notes and a small head nod on every beat. Her hair moves slightly, as if in a light wind. The camera holds a steady medium close-up. Keep her face, earring and jumpsuit unchanged.

How do you cut AI shots on the beat?

1. Lock the master on the timeline first. Shots are trimmed to the music, never the reverse.
2. Mark the beats with beat detection or a marker on every downbeat.
3. Cut on downbeats; change shot size on bar lines. A cut a few frames early reads as an accident.
4. Hold sung shots for a full line.
5. Check p, b and m. The lips must close on those frames. Slide a late take a frame or two; re-sync a drifting one.
6. Mark each chorus with a change of location, light or shot size on its first beat.

For reframing and pacing in vertical edits, see our vertical video editing guide.

What rights and consent do you need?

A commercial song carries two sets of rights, the composition (songwriters and publishers) and the recording (usually a label), and a music video needs both unless the song is yours.

YouTube lists AI-generated music as something creators must disclose (the "AI use" setting under Attributes in Studio) and says disclosure won't limit audience or eligibility to earn. Instagram names "a song created using AI-generated vocals" as needing its AI label, and TikTok requires a label on AI content with realistic images, audio or video. A commercial-use grant lets you use a track; whether a mostly machine-made song can be copyrighted is a separate question. Not legal advice; see using AI video commercially and whether AI videos can be monetized.

How do you cut it down for Shorts, TikTok and Reels?

Most listeners meet the song in a cut-down, so build it from the hook.

1. Open on the chorus or the strongest line; the first second should already be singing.
2. Reframe for 9:16. If you generated 16:9, keep the singer's face in the centre third, about all a vertical crop keeps.
3. Burn in the lyrics and proofread them: sung words over a full mix are harder to transcribe than speech.
4. Watch the clock on YouTube. Shorts can run three minutes, but one over a minute with an active Content ID claim of any type is blocked and can't earn. If the track is in Content ID (your own release can be, once a distributor registers it), keep Shorts to a minute.
5. Make several cut-downs, post them over days, and label realistic AI on each platform.

Where does ClipSpeed fit?

For the generated shots, ClipSpeed's AI Creator puts the pieces in one workspace: Nano Banana Pro, GPT Image 2.5 Flare and GPT Image 2.5 Sunburst for reference sheets and start frames, Seedance 2.5 and Veo 3.1 for performance shots where the generated vocal is the point, Kling 3.0 Turbo and MiniMax H3 Max for cheap, fast b-roll, and Gemini Omni 1.1 Flash for editing a take by instruction. ClipSpeed Create, in early access, brings generation into Claude or ChatGPT (how it connects); each generation is quoted first and runs only after the account owner approves it on a signed-in ClipSpeed page.

The cut-down is ClipSpeed's other product, AI clipping: paste the URL of a finished video or livestream and it finds the strongest moments, cuts them to 9:16 or 16:9, burns in captions and gives each clip a viral score. See also clipping music and DJ content.

Which approach fits your project?

• A concept video for an AI-made song → performance shots with the lyrics in the prompt.
• Your real vocal on an invented singer → an audio-driven portrait driven by the vocal stem, intercut with b-roll.
• A staged, moving performance that must match the master → generate the shot, then re-sync it.
• Cheap b-roll in volume → Veo 3.1 Fast or Lite at 720p, muted under the master.
• No song yet → a music model whose terms grant commercial rights.

The bottom line

Lock the song, plan in bars and give every section one job. Decide early whether the voice on screen must be your recording; if so, drive a portrait with the vocal or re-sync the footage. Spend on the chorus, cut on downbeats, check every p, b and m, and clear both sets of rights before anything goes public.

Sources and method

Every limit and price above was read on October 8, 2026 on the provider's own docs, model page or pricing page. Thirty-second costs are list rate times 30 seconds for one attempt. No generations were run for this guide; the song and the singer are invented. Not legal advice.

• Dreamina Seedance 2.5 tutorial and prompt guide (BytePlus ModelArk)
• ModelArk model pricing (BytePlus)
• Generate videos with Veo 3.1 and Gemini Developer API pricing (Google)
• Generate music with Lyria (Google)
• Kling VIDEO 3.0 model user guide (Kling AI)
• OmniHuman 1.5, Kling AI Avatar v2 Pro, Standard and Kling LipSync (fal)
• OmniHuman-1.5 project page (ByteDance)
• Lipsync models and pricing (sync.so)
• Suno pricing
• Understand three-minute YouTube Shorts, Disclosing use of GenAI content and Our approach to responsible AI innovation (YouTube)
• Label AI content on Instagram (Instagram Help Center)
• About AI-generated content (TikTok Support)

Related guides

GuideNative Audio in AI Video: Sound, Lip-Sync, Music GuideConsistent AI Characters Across Video Shots RightsCan You Use AI-Generated Video Commercially?

• How to Clip Music and DJ Content into Viral Social Media Posts
• Can AI-Generated Videos Be Monetized? What YouTube, TikTok and Meta Actually Check