Music Video
Juno Grey

Third Person

A Juno Grey music video where the AI finally sings on camera. Two rooms, 22 clips, and a lip-sync model fed the exact seconds of vocal it had to perform.

3:38
Runtime
22
Clips in the final cut
6
Cuts to reach final
687
Higgsfield credits
Part of the Juno Grey project
Video three for an AI pop star built from a spreadsheet
The whole video on one line

Two rooms, 22 shots

0:003:38
Candlelit dinner. Talked about, never to.White room. Sings straight down the lens.Hallway between them.
14 sung by Wan 2.7 from the real vocal 8 silent, Kling 3.0
  • Shot 1: Dinner dolly0:00
    Dinner dolly
  • Shot 2: Hand on chair0:09
    Hand on chair
  • Shot 3: Edge of the photo0:20
    Edge of the photo
  • Shot 4: Hallway0:27
    Hallway12s
  • Shot 5: Chorus 10:39
    Chorus 110s
  • Shot 6: Lean in0:49
    Lean in12s
  • Shot 7: Stands up1:01
    Stands up
  • Shot 8: Out the door1:09
    Out the door
  • Shot 9: Windowsill1:17
    Windowsill12s
  • Shot 10: Mirror1:29
    Mirror12s
  • Shot 11: Whisper1:41
    Whisper13s
  • Shot 12: Chorus 21:54
    Chorus 213s
  • Shot 13: Sway2:07
    Sway15s
  • Shot 14: Empty chair2:22
    Empty chair
  • Shot 15: Bridge ECU2:25
    Bridge ECU14s
  • Shot 16: Hold, 0.25x2:39
    Hold, 0.25x11s
  • Shot 17: “Talking now”2:43
    “Talking now”7s
  • Shot 18: Final chorus2:50
    Final chorus14s
  • Shot 19: Doorway look-back3:04
    Doorway look-back14s
  • Shot 20: Last line3:18
    Last line10s
  • Shot 21: Morning after3:27
    Morning after
  • Shot 22: Pull-back3:31
    Pull-back
How it was made

Words → voice → stills → motion → cut

Lyrics
Claude
4 drafts, then approved
Song
Suno
152 BPM, G♯ minor
Map it
whisper + librosa
every lyric gets a timestamp
Keyframes
Nano Banana Pro
36 stills, 2 identity refs
Sung shots
Wan 2.7
keyframe + exact audio slice
Silent shots
Kling 3.0
8s, sound off
QC
mouth strips
sync offset must be 0.000s
Assemble
ffmpeg
one timeline table, 22 renders

The trick is the fork: every sung shot gets the exact seconds of vocal it covers, so the lip sync is real, not guessed.

The models

Cast & crew

  • Claude Fable 5.1
    Director, lyricist, editor, QC
    10
    pipeline steps
  • Suno
    The voice. Only model that sings.
    1
    master track
  • Nano Banana Pro
    Every still frame
    36
    keyframes
  • Wan 2.7
    Lip sync from real audio
    27
    generations
  • Kling 3.0
    Wordless cutaways
    11
    clips
  • whisper + librosa
    Timestamps, tempo, sync check
    0.000s
    sync offset
  • ffmpeg 6
    Slices, renders, concat, mux
    22
    segments
  • Not castSeedance 2.5, Grok Video 1.5, MiniMax: 25 to 65 credits a clip. Too rich for 22 shots.
Quality control

Does her mouth match the master?

Eight mouth-crop frames from the chorus 1 clip at half-second intervals, used to check lip sync
0.5s
between frames, cropped to the mouth, for every Wan clip
0.000s
offset when each clip's audio is cross-correlated against the master. All 14 passed.
2 retakes
for slices that started on instrumental. She mouthed early, then sat closed on the line.
Stills first

The contact sheet

Contact sheet of 16 Nano Banana Pro keyframes: the dinner table, hallway, white room, windowsill, mirror, and doorway
Every clip starts from one of these. Two identity references keep the face, hair, and skin consistent from shot to shot. Spot the wrist tattoo in row three. That is why v2 exists.
Iteration

Six cuts

  1. v1
    Everything built
  2. v2
    Tattoo gone, 10 shots reframed
  3. v3
    Skin fixes
  4. v4
    New ending, new opening
  5. v5
    She walks out
  6. v6
    Bridge jump cut killed
Opening shot
Opening shot, v1 · head of the table, wideBEFORE
v1 · head of the table, wide
Opening shot, v6 · side of the table, medium, sleek hairAFTER
v6 · side of the table, medium, sleek hair
Bridge at 2:39
Bridge at 2:39, v3 · jump cut into a second close-upBEFORE
v3 · jump cut into a second close-up
Bridge at 2:39, v6 · re-angled to the doorway, re-slicedAFTER
v6 · re-angled to the doorway, re-sliced
Higgsfield credits

What three videos cost

Nano Banana Pro stillsWan 2.7 lip syncKling 3.0 clips
Sleep Mode
shelved
168
Home Game
shelved
184
Third Person
released
687
1,039
credits, three videos
1.5
credits per second of lip sync
2
credits per still
Now in the Juno brand kit

Six rules, learned the expensive way

  • Lyrics first, always

    Nothing renders until the words are approved.

  • Juno has no tattoos

    Prompt clean skin. Use a clean frame as the second identity ref.

  • Medium or tighter

    Wan loses the mouth when the face is small.

  • Start slices on the vocal

    Instrumental lead-in makes her mouth early. Cover gaps with a hold.

  • 0.000s or it's wrong

    Cross-correlate every clip's audio against the master.

  • Keep prompts PG

    “Bites her lip” + slip dress trips the safety filter.

Tools Used

CLClaude Fable 5.1SUSunoNANano Banana ProWAWan 2.7KLKling 3.0HIHiggsfieldFAfaster-whisperLIlibrosaFFffmpegPyPythonVIVimeo

Key Topics

AI Music VideoLip SyncImage-to-VideoKeyframe GenerationAudio AnalysisCharacter ConsistencyffmpegHiggsfieldSuno

More Like This

Juno Grey thumbnail
Spotify

I Created a Pop Star

A fully AI-generated pop star created from a data model that ranked 300+ real artists across four dimensions of cultural staying power.

AIMusicData
View project
The Pixar Trailer I Made For Myself (Because I Could?) thumbnail
VimeoSubstack

The Pixar Trailer I Made For Myself (Because I Could?)

An extremely unnecessary experiment in AI filmmaking.

AI VideoCreative Experiment
View project
Shofar So Good thumbnail
Substack

Shofar So Good

An alternate reality where Kanye apologizes — told in 6 AI-generated tracks.

MusicAISatireEP
View project
← Back to all Create projects