>_ chinchilla.marketing Start a project
← all video builds ~/builds $ ./lyrics-caption-build.py --render

Lyrics, synced. Giant captions.

A 10-second lyrics-caption video build. Halftone coins animate as a mirage spin with a stronger twinkle while the picture and subject stay still — and ElevenLabs Scribe transcriptions render as giant, beat-synced background captions, one word across the canvas, three words stacked, each example using a different overlay blend.

ElevenLabs Scribeblend overlays30fps10s loop
lyrics-caption-build — sample output · 10s
● autoplay · muted · seamless loop — the exact file you receive the recipe for 1080×1080 · 30fps
Unlock free — 32s sponsor view → See the GIF
examples

The loop, as a GIF.

A share-ready GIF of the lyrics caption build — rounded corners, infinite loop. Drop it in a chat, a deck, a story, or a carousel and it just runs.

Lyrics Caption build — animated GIF example
lyrics-caption-build.gif · rounded · loops forever · 1080×1080 · 30fps
what this build is

Karaoke that reads itself.

The caption is the visual. Every word lands on the beat, drawn huge across the frame so it works even with the sound off.

01

Word-synced captions

ElevenLabs Scribe turns the vocal window into word timings; the script draws each word in sync with the audio.

02

Mirage spin halftones

The halftone coin lattice spins and twinkles in the background while the source photo stays perfectly still.

03

Per-example blend modes

soft-light, multiply, and overlay change how the giant captions sit in the mix — one look per example.

04

Studio strobe cycle

A 2.5 Hz background flash cycles through screen / add / difference / add accents for extra punch.

how it works

From song to synced loop.

A deterministic four-step pipeline. No generative model — every frame is drawn by code, so it's reproducible and re-skinable.

01 pick
Choose a song and a 10-second window.
02 transcribe
Run the window through ElevenLabs Scribe → word-level timings.
03 render
Draw 30fps frames: halftone coins spin + twinkle, captions render word-by-word in sync.
04 mux
ffmpeg muxes the song back in → lyrics_1.mp4 (1080×1080, H.264 + AAC).
what's inside

The .md build guide.

You're getting a downloadable .md build file written for AI agents (and humans). It contains the full recipe, the exact dependencies, the run command, and every knob to tune — plus the finished sample video is embedded above so you can see exactly what it produces.

lyrics-caption-build.md
the downloadable guide — full recipe, dependencies, run command, every knob (this is the file you buy)
animate_lyrics.py
the build script (the recipe, in code)
lyrics_1.mp4
finished sample output — 10s, synced audio (shown above)
neon_arrangements.py · neon_halftone.py · pfp_halftone_subject.py
shared halftone glyph / icon / intensity-map helpers
output
1080×1080 · 30fps · 10s · H.264 + AAC · seamless-loop friendly
typeface
JetBrains Mono ExtraBold (same as chinchilla.marketing)
caption sync
ElevenLabs Scribe word timings
overlay blends
soft-light / multiply / overlay · lyric blends screen / difference / add
grain
gaussian σ36 + fine noise 16
dependencies
Python 3.10+, numpy, Pillow, ffmpeg on PATH, ElevenLabs API key
compare & contrast

Chinchilla build vs. Higgsfield AI models.

Our builds are deterministic code — the same recipe renders the same frame every time. Higgsfield's lineup (Seedance 2.0, Kling 3.0, Veo 3.1, Soul, Marketing Studio) is generative AI — diffusion and transformer models that sample new pixels each run. Different tools, different jobs.

dimension
this build
higgsfield ai
How it renders
Code-authored. numpy + Pillow + ffmpeg draw every frame. Same input → same output, version-controllable.
Diffusion / transformer models sample new pixels each run — no two renders are identical.
Control
Per-pixel. Edit a constant (wave speed, grain, blend mode) and re-render the exact change.
Prompt + seed steering. Interpretive; you guide the model rather than specify frames.
Identity & subject
Your actual asset. The real photo / subject cutout stays pixel-faithful; zero identity drift.
Soul Characters and reference images keep identity, but generative output can drift; or it invents people who don't exist.
Brand lock
Baked in. Fonts, palette, layout and motion are exact brand tokens in code.
Brand look is approximated via prompt; Marketing Studio adds product/avatar retrieval for ads.
Cost model
Free to unlock — a short sponsor view sized to the render workload releases the recipe, then free local renders forever.
Per-generation credits / API cost on every video you produce.
Speed
Seconds to minutes locally, batchable, no queue.
API job queue — minutes per render, rate-limited by account.
Output style
Graphic motion. Halftone, typographic and blend-mode loops tuned to music.
Photoreal, cinematic, or stylized AI footage — scenes, actors, b-roll.
Best for
On-brand loops you re-run daily — lyrics / karaoke, quote-of-the-day, drop announcements, and caption-first socials from a real photo.
Footage that doesn't exist, actors & avatars, ads with people, net-new b-roll.

They're complementary. Generate the hero asset with Higgsfield, then feed it into a Chinchilla build to make it an on-brand, loopable, music-synced social video you can re-run every day.

Get the Lyrics Caption build.

Free download. Unlock the lyrics-caption-build.md guide with a short sponsor view — the same recipe that renders the video above. Paste it into any AI agent and it can rebuild, remix, and re-skin the effect for your own content.

downloadable .md for AI agents instant delivery
free no account · unlocked by a sponsor view
Unlock the .md →
32s watch time — scaled to this build's render workload (700 units), capped at 60s