No Video Model: Claude Opus 5.5 Coded a 50-Shot Painted Music Video Entirely in Code
DigitalDaydreamers1 · reddit · 2026-09-25
A Redditor gave Claude Opus 5.5 a Suno-generated MP3 and 15 short art-direction messages; the model built a 50-shot painted music video entirely in code, with no image or video model involved. It analyzed the audio (beats, tempo, 32-band spectrum), pulled lyrics from MP3 metadata, installed an offline speech aligner to sync story beats to vocals, adapted an MIT-licensed WebGL brushstroke renderer, and drew every character and set procedurally. It ran a headless browser in its sandbox to render test frames and self-corrected blown-out glows, vanishing cars, awkward poses, and crashes unprompted. Final render: 4,698 frames at 1080p30 in software over 2+ hours, resuming through two sandbox restarts. Limitations: stylized look built on an existing renderer, needs manual "continue" prompts over long renders, and lyric timing is imperfect. A year ago this was a week of specialized work; now it's a conversation.
More from coding & agent
- OpenRouter launches Server Tools Marketplace: web search, shell, image gen and model calling — jeff_weinstein · 2026-09-25
- With AI agents handling calendars, is Calendly-style p2p scheduling already dead? — heyneighbor · 2026-09-25
- Visualizing 2973 agents that built vgpu: custom multi-agent pipeline with Vercel AI SDK — evilrabbit_ · 2026-09-25
- NT kernel engineer: Linux's messy permission model is a poor fit for AI agents — lauriewired · 2026-09-25
- Going local-first: 2x 5060 Ti at ~100 tok/s plus a 96GB M5 Ultra, betting on a 3-year inference floor — No-Name-Person111 · 2026-09-25
- Founder: ChatGPT Astra replicated 3 years of framing-solver work in SketchUp in 10 minutes — natesiggard · 2026-09-25