🔧 Most “AI video generators” output slop with drifting frames. I’m a dev — I wanted deterministic, frame-perfect 3D. So I built my own engine. Here’s the architecture (with a teardown of an actual engine it generated 👇).
TL;DR: I built Unstill — you give it a prompt, an LLM authors a pure-data scene, a custom Three.js + GSAP engine renders it interactively, and Remotion exports a narrated MP4. The hard part wasn’t the LLM — it was making the live preview and the exported video pixel-identical.
How it works:
<ol>
<li>The LLM doesn’t render — it authors data. It outputs a file: pure data describing objects, tweens, camera moves, and chapters. No code-gen, no pixels — so output is validatable, repairable, and version-controllable.</li>
<li>One engine, two outputs. A Three.js + GSAP + Matter.js engine renders that data for BOTH the live studio preview and the video export. The catch: “looks right in preview, broken in the MP4.” The fix was making all motion a pure function of timeline.time() (no accumulated dt), so any frame renders identically and Remotion can parallelize the render across workers with zero drift.</li>
<li>Remotion for the MP4 — with AI narration (TTS) synced per chapter, captions, and a coded like/subscribe/bell outro shared between preview and export so they never diverge.</li>
<li>Quality is a loop, not a prompt. Structured-output schemas + domain accuracy packs + a visual-critique pass that re-renders and self-corrects.</li>
</ol>
Here’s a 4-stroke engine it generated end-to-end 👇
▶️ <a href="https://www.youtube.com/watch?v=TWUhj6Hmv3M" target="_blank" rel="noopener nofollow">https://www.youtube.com/watch?v=TWUhj6Hmv3M</a>
What’s your thought on this?
#threejs #webgl #javascript #typescript #react #nextjs #ai #machinelearning #webdev #devtools

Varun Dev

A developer built Unstill, a deterministic 3D video generation pipeline to avoid the frame-drift issues common in AI video generators. The architecture separates concerns: an LLM authors a pure-data scene file (no code generation), a unified Three.js + GSAP + Matter.js engine renders it for both live preview and video export, and Remotion handles MP4 export with TTS narration synced per chapter. The key technical insight is making all motion a pure function of timeline.time() rather than accumulated delta-time, enabling Remotion to parallelize rendering across workers with zero drift and ensuring pixel-identical output between preview and export. Quality is improved through structured-output schemas, domain accuracy packs, and a visual-critique self-correction loop.

Most AI video tools output slop. I built a deterministic Three.js + LLM pipeline instead