Guide · Visuals
Lyric video instead of music video: syncing text without a studio
A music video with actors is not an option for an anonymous AI project. A lyric video is – it is the format YouTube favours for faceless music anyway, and it turns an album into a 40-minute video listeners play through in one sitting. Our first attempt was boring; the second one works. This is the second one.
What made the first attempt boring
One motif per song, looped forward and backward, text on top. Technically clean, tiring after 90 seconds. The lesson: a lyric video needs editing. Not action, but change – a new image roughly every four bars.
Motif stock: eleven images per song
Around ten motifs per song, all from the song’s world but never showing the same thing: jacket over the chair, shoes by the door, skyline through the window, rain on the glass. About half as moving clips (image-to-video), the other half as stills with a slow camera move – those cost seconds of compute and stand up fine next to the clips. Moving motifs go in the choruses, calm ones in verses and bridge.
An edit grid from the song
- Determine the song’s tempo (most audio tools estimate BPM automatically) and derive the bar length.
- Mark the sections: intro, verse, pre-chorus, chorus, bridge, outro – rounded to bar boundaries.
- Split each section into blocks of at most four bars; each block gets a different motif from the section’s rotation.
- Run shots forward only, never ping-pong; transitions as short crossfades (0.6 s).
Result for a four-minute song: around 30 shots. That sounds like a lot of work, but it is a table – and the table is reused for the next song.
Setting the text: the rules that proved themselves
- Always two lines, never across a verse boundary. The second line appears about 0.4 seconds after the first – it feels spoken, not stamped.
- Chorus different from verse: larger and in the brand colour; verses smaller in a lighter tone. That way the song’s form is visible without knowing it.
- Readable over bright images: no hard drop shadow but a soft, opaque outline around the letters. Drop shadows vanish on skylines; the outline stays.
- Lower third instead of a full-screen card: number, line, title at bottom left when the song changes. A full-screen card interrupts; a lower third accompanies.
- Brand name small, permanently bottom right. Not as a logo, as a wordmark.
Timestamps without word-accurate transcription
The hurdle in a lyric video is timing: when does which line appear? Speech recognition gives segment times, but for sung text often the wrong words. The approach that works: use recognition only for the times, not the text. Estimate word times evenly within a segment (sung, it is about two to three words per second) and align that estimate against the real lyrics you wrote yourself. That is accurate to about half a second – entirely sufficient for two-line overlays. Nobody needs word-accurate karaoke effects.
Detecting the chorus automatically
A line that appears at least three times in the lyrics is chorus. That simple rule is enough to set choruses large and gold and give them the moving motifs – without marking every song by hand.
What else we learned
- Image-to-video models like to turn one moving car into two. Solve moving vehicles as stills with a camera move.
- In bar and club motifs a hand often reaches into frame towards the end. Cut clips before the end or put hands on the negative list.
- For a whole album, render shots individually and assemble at the end – aborted render runs otherwise leave unusable files that only show up during assembly.
- Chapter markers in the video description: a timestamp per song. YouTube shows them as chapters, and listeners jump straight to their favourite.
A lyric video is typography with a background. Set the type right and you get very far with very little image.
All guides are deliberately kept general: which models and tools run on your machine depends on your hardware. Questions or corrections? Write to us.