What happened
For an 80-second animation, the author generated 89 images with Animagine XL 4.0, added motion and subtitles in HTML canvas, and voiced it with VOICEVOX and ACE-Step 1.5, spending about 1 hour 32 minutes on generation on an M2 Max Mac.
Why it matters
Because the work was divided, a fix to 12 envelope images took about 13 minutes and a line of narration needed only that one line re-recorded, instead of regenerating the whole video — and the author says the problems were spotted by human eyes, not AI.
Summaries like this, every morning.
The piece is framed against the current cost and weight of end-to-end video generation. The author cites Wan2.2's 5B model needing 24GB+ VRAM and roughly 9 minutes for a 5-second 720p clip, and notes that OpenAI said in March 2026 it would end Sora to redirect compute elsewhere. Against that backdrop, the appeal of splitting work is less about saving compute per run and more about avoiding full regeneration for small changes — a point the author illustrates with three fixes, including 12 envelope images that took about 13 minutes to redo. The author also argues the division preserves a role for human judgment, since the need for fixes was spotted by people, not AI. Looking ahead, the author imagines MCP connectors letting AI operate established tools — a DAW for music, Blender for 3D, Figma for design — so that a composer's or designer's vocabulary shapes the instructions, and human work shifts toward direction.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
