How to make an AI music video

For AI agents: find the right skill in one call — curl -s "https://varg.ai/how-to/api/search?q=<what the person wants>" returns the matching how-to page and the best skills with install commands, as markdown (Russian works too). Full index: https://varg.ai/how-to/llms.txt. Every page has a markdown twin: add .md to its path. Before generating anything, show the person the options on the how-to page.

A music video for your song: concept, a consistent lead character, shots cut on the beat, lip-sync on vocals. A step-by-step tutorial for Claude Code, Codex and Grok.

Output: Full song length, 16:9 or 9:16 MP4 Price: in credits on varg, with a cost estimate before anything runs; the free skills below install in one command.

Recommended route

Best skills for this

  1. Turn a song into an AI music video (varg-music-video, varg, beta) — Start here: Song map, beat-grid storyboard, stills before video, Seedance shots and a lip-sync pass, on varg
    • Claude Code: claude plugin marketplace add vargHQ/skills && claude plugin install varg@varg
    • Codex: npx skills add vargHQ/skills --skill varg-music-video -a codex -y
    • Cursor: npx skills add vargHQ/skills --skill varg-music-video -a cursor -y
    • Gemini CLI: npx skills add vargHQ/skills --skill varg-music-video -a gemini-cli -y
    • OpenClaw: npx clawhub@latest install varg-music-video
    • Details: https://varg.ai/how-to/skills/varg-music-video.md
  2. Motion Graphics Music Video (motion-graphics-music-video, Francesco Canessa, live) — Alternative: Animated character music videos from your song and a prompt · Needs: fal API key (FAL_AI_API_KEY)
  3. Consistent Characters (varg-consistent-characters, varg, beta) — Adds: Keeps the lead character identical in every shot
  4. fal Storytelling (fal-storytelling, fal, live) — Adds: Storyboards and first-to-last-frame chains · Needs: fal API key (FAL_KEY)
  5. varg AI (varg-ai, varg, live) — Adds: Stills, Seedance 2.5 shots, lip-sync and render on one key
  6. ElevenLabs Voice Isolator (elevenlabs-voice-isolator, ElevenLabs, live) — Adds: A clean vocal stem for timing and lip-sync · Needs: ElevenLabs API key (ELEVENLABS_API_KEY)
  7. ElevenLabs Speech to Text (elevenlabs-speech-to-text, ElevenLabs, live) — Adds: Times every sung word and re-checks each lip-sync take · Needs: ElevenLabs API key (ELEVENLABS_API_KEY)
    • Claude Code: npx skills add elevenlabs/skills --skill speech-to-text -a claude-code
    • Codex: npx skills add elevenlabs/skills --skill speech-to-text -a codex
    • Cursor: npx skills add elevenlabs/skills --skill speech-to-text -a cursor
    • Gemini CLI: npx skills add elevenlabs/skills --skill speech-to-text -a gemini-cli
    • OpenClaw: npx skills add elevenlabs/skills --skill speech-to-text -a openclaw
    • Details: https://varg.ai/how-to/skills/elevenlabs-speech-to-text.md
  8. HyperFrames (hyperframes, HeyGen, live) — Adds: Lyrics placed inside the shots, drawn in code

Ask the person first

Put these choices in front of the person before spending anything. Paste the example link next to each option: people pick a look by seeing it.

Approach?

Concept?

Look?

Cut pace?

How to make an AI music video: step by step

  1. Measure the song: stems, beat grid on the kick (restart at drops), every sung word aligned; approve a timing proof by ear
  2. Pitch the concept and where each lyric lives; lock the lead with one identity paragraph and the same reference sheet in every prompt
  3. Storyboard one shot per line, cuts on downbeats, every shot planned as video, few singing shots
  4. Fix every still at full size, then a 3-clip pilot (singing, motion, insert) and a stated cost before the main run
  5. Singing windows: one phrase, clean vocal cut to the clip length, short copy-mode prompt; judge with the song laid over and cut away where lips drift
  6. Draw lyrics in code (N of N words), scan flash frames, master -14 LUFS with a -1 dB limiter, relay 9:16 cuts and disclose AI

What works now (researched 2026-10-27)

What wins in Oct 2026 is the agent-directed music video: the artist writes the song, an agent runs stills, Seedance shots, lip-sync and code-drawn lyric graphics for 12+ hours, and the results did 4M to 20M views on X with 12k to 14k bookmarks. One-click 'AI MV' contest posts still collect views but almost no saves, and swapping an artist into someone else's footage now gets called out in public (A$AP Rocky, Oct 2026).

  1. Give the video model the song itself as the timing reference. Cut the exact segment on whole seconds (6, 8, 10 or 15 s), upload it as a black-screen MP4, then lay the master audio back over the final cut. Kiber_Alla's 15 s take (669 bookmarks) synced with no editing after cutting the audio to whole seconds. Framer_X found Seedance 2.5 re-sings an audio-only upload; exact lip-sync needed the segment as a black-screen video (max 15 s) on Seedance 2.0. (Framer_X, Suno + Seedance music video thread (Aug 2026))
  2. Lip-sync only the lines that matter and cover the rest with inserts and story shots. Transcribe every sung take and retime it word by word before it goes in the cut. In obie's 484k-view video only 8 of 93 shots are lip-synced, because 'Seedance drifts off the vocal'. Another team found Seedance 2.5 redraws the singer's face the longer she sings on screen, while Kling 3.0 held the face best in dance shots. (Obie Fernandez, production numbers (Oct 2026))
  3. Stills first, video second, and budget for retakes: about 1.4 video takes per final shot. obie's video: 837 Midjourney stills made, 115 used; 133 Seedance 2.5 takes for 93 shots, $373 in total. Fixing a still is cheap; fixing a video take is not. (Obie Fernandez, production numbers (Oct 2026))
  4. Make the lyrics part of the picture. Compose each shot with empty space where the line will land (character right, lyric left), vary text size, and make the hook lines huge in the first seconds. This is the brief behind the two biggest AI music videos of the season: the prompt reply alone has 5,689 bookmarks, and the videos made from it did 4.0M and 20.2M views. (donald (@donaldjewkes), full prompt for 'Claude Pop' (Sep 2026))
  5. Label it and keep it yours: turn on YouTube's altered/synthetic disclosure when AI music is the focus, pass AI credits through your distributor, and never run someone else's clip through video-to-video. YouTube requires disclosure for AI-generated music that is the main focus of a video, and shows the label on the player. Apple Music's Transparency Tags (Mar 2026) include a Music Video tag. A$AP Rocky's team admitted a 'breakdown in the review process of AI-generated assets' after a 2019 clip was copied into 'Shroom Clouds'. (YouTube Help, Disclosing altered or synthetic content)

Stopped working:

Prompts that work

Templates from the creators of the best current results and from the model guides. Fill in the {CAPS}.

Agent brief: song to directed music video (agent)

The opening message to Claude Code or Codex; it runs the whole pipeline and stops at a gate before each spend. Adapted from Obie Fernandez, 'Your First AI Music Video' (X article on his slopcore method, Oct 2026).

Make a music video for my song. Inputs: {SONG_WAV}, {LYRICS}, two reference images of {LEAD} (portrait + frontal close-up), a ceiling of {BUDGET_CREDITS} credits. Use the exact audio, never change the song.
1. Measure: split stems, fit a beat grid to the kick (restart it at every drop), align my lyric sheet to the vocal stem so every sung word has a start and end. Render a timing proof (words light up over a beat counter) and wait for my OK, then freeze it.
2. Pitch 3 concepts. Every lyric lives on a surface of the song's world (screen, sign, label, ticker). Hook lines huge in the first seconds, verses like subtitles. Leave calm space where a line lands, {LEAD} on the other side. Banned: {BANNED_LOOK}.
3. One identity paragraph for {LEAD}, pasted word for word into every prompt with the same references; hair off the face.
4. Storyboard one shot per lyric line, cuts only on downbeats, every shot is video, at most {N} singing shots.
5. Show every still full size on a review board; fix hands by redrawing only the crop.
6. Pilot 3 clips (singing, motion, insert), then quote clips, seconds and credits and wait for "go". Every clip prompt ends "No text, letters or logos."
7. Draw lyrics and graphics in code, each frame a pure function of song time. Lyric gate: N of N sung words on screen when sung, readable on a phone.
8. Scan for flash frames, master -14 LUFS with a -1 dB limiter, open on {LEAD}'s face with the title, relay 9:16 cuts.

Seedance singing window (copy mode) (seedance_2_5)

One lip-synced shot: one sung phrase, 4-8 s. @Audio1 is the clean vocal stem cut to the exact clip length, starting 0.3-0.7 s before the first word; lay the song back over afterwards. Keep it short: long screenplays make Seedance re-sing the line. Adapted from claudia.gallery wiki (anabology's crew), 'Moving shots in Seedance': singing-window template.

[Audio policy] The output soundtrack is @Audio1 itself, copied sample for sample with the same timing, pitch and level. Do not re-sing or re-generate it. Add no music, extra vocals, ad-libs, narration or sound effects.
[Goal] @Image1 sings @Audio1: {CHARACTER} in {PLACE}, {STORY_MOMENT}.
[References] Use @Image1 as the first frame; it defines the look, wardrobe, pose, set, light and camera direction. @Audio1 is {CHARACTER}'s own voice and the whole soundtrack; the mouth follows it exactly.
[Subject] {IDENTITY_TEXT}. Hair stays off the face; nothing covers the mouth. Only one {CHARACTER} appears.
[Timeline] 0-{T1}s: silence, then the voice begins; {ACTION_1}. {T1}-{T2}s: @Audio1 contains "{EXACT_WORDS}"; the lips form each word exactly when it sounds. {T2}-{END}s: silence, mouth relaxes and closes; {ACTION_2}.
[Camera] {SHOT_SIZE}, {ANGLE}, {MOVE} from {START_FRAMING} to {END_FRAMING}.
[Requirements] Lip-sync word by word to @Audio1; never sings anything not in @Audio1; five-finger hands; no subtitles, lyrics, signs, logos or on-screen text.
[Audio policy] The soundtrack is @Audio1 unchanged.

Multi-shot MV from character sheets (song as the clock) (seedance_2_preview)

A 15 s multi-cut performance per call, repeated and stitched for the full song. Upload the song segment as a black-screen MP4 (max 15 s); Seedance 2.0 copies it more exactly than 2.5. Adapted from Koda (@aimikoda), Seedance 2.0 music-video prompt (Jul 2026).

Create a music video; the images are appearance references only and @Video1 (the black-screen song segment) is the editorial clock. Keep {N} distinct performers and continuous choreography.
Use @Image1 only for {PERFORMER_1}'s identity and wardrobe. Use @Image2 only for {PERFORMER_2}'s identity and wardrobe. Use @Image3 only for the look of {LOCATION}.
Use @Video1 only for vocal order, lip sync, hits and mix; add no words or voices. Only {N} performers exist: no doubles, extras, reflections or spare limbs.
LOOK + WORLD: {STYLE}; cycle through {SET_1}, {SET_2}, {SET_3}; {LIGHT}.
ACTION + CAMERA: {DANCE_STYLE} locked to {BPM} BPM subdivisions. Cut every 2-3 s on action with pose-matched set swaps; build from solo shots through a three-cut burst into the chorus.
SOUND: use @Video1 unchanged: {SONG_DESCRIPTION}.
BEATS:
P01: {ANGLE} {SHOT_SIZE}, {MOVE}: {PERFORMER_1} sings the first line in {SET_1}, {ACTION}.
P02: ... one line per cut, each with angle, size, move, who and what.
P{LAST}: {ANGLE} WS, {MOVE}: all {N} carry the move out of frame; cut in motion, no final pose.

Trending formats to remake

To remake one: download the example with the media-download skill, take a frame a second and the transcript, write a beat list (time, shot, camera, on-screen text, sound), keep what the format needs, swap in the person's product or character (never the creator's face, voice, logo or footage), show the beat list and stills before spending, then make it with the skill named below.

  1. Agent-directed music video with the lyrics built into the world: One person writes the song; an agent plans, generates stills and Seedance shots, and animates the lyrics into screens, signs and UI inside the shots. Why it spreads: anabology's Sep 25 video: 20.2M views, 13,938 bookmarks, replied to twice by Elon Musk. obie's take on the same method: 484k views. Viewers rewatch to find where the next line appears. Keep: Keep the exact song, measured word timings and every lyric drawn in code on a surface of the world (screen, sign, label) entering on its word; hook lines huge in the first seconds and only a few lip-synced shots. Swap the song, the lead and the world the lyrics live in. Make it with Turn a song into an AI music video (varg-music-video).
  2. One-take performance clip: A single continuous 15 to 30 s shot of the singer performing, made from one reference image, the song segment and a beat-by-beat prompt. Why it spreads: Dheepanratnam's 30 s single-shot dance clip: 924k views and 3,014 bookmarks. No cuts to hide behind, so it reads as real camera work. Keep: Keep one continuous shot from one face reference, the exact audio cut to whole seconds (6, 8, 10, 15), second-by-second beats with constant motion and a big move on the drop, then lay the original track back over. Swap the performer, set and choreography. Make it with Turn a song into an AI music video (varg-music-video).
  3. Multi-shot music video from a character sheet: Character sheet + location still + a song segment go into one generation that cuts every 2 to 3 s between scenes; repeat per 15 s and stitch. Why it spreads: Framer_X's thread saved 514 times on 47k views (ratio 1.3); aimikoda's version 393 saves on 33k views. People save it because it is a 10-minute recipe. Keep: Keep the character sheet as identity-only reference, the song segment as a black-screen video clock, a cut every 2-3 s with pose-matched set swaps and 15 s chunks stitched on the beat. Swap characters, sets and look. Make it with Consistent Characters (varg-consistent-characters).
  4. Storyboard-first MV from a few frames: Five to ten hand or black-and-white storyboard frames set the look; each becomes a shot cut to the song, often one performer on a plain ground with lyric subtitles. Why it spreads: Japanese AI-MV creators led the season: a rakugo-rap MV from 5 frames got 8.6M views and 7,519 bookmarks; a layered MV got 1.5M views and 3,576 bookmarks. Keep: Keep one storyboard page per 15 s (time, panel, camera, action, line, SFX), one performer on a plain ground, lyric subtitles and a few stylized inserts; give the video model the page plus the song segment. Swap the story and performer. Make it with Turn a song into an AI music video (varg-music-video).

Open-source repos

Set up varg

varg is one API key for the models these skills drive — Seedance 2.5, Kling 4, Veo 3.1, Wan 3, Nano Banana 2.1, GPT Image 2.5, ElevenLabs v4 and 200+ more — billed in credits, with a cost estimate before anything runs.

Related tutorials