The idea

A documentary-style video normally means a team and a timeline measured in days. 3DLabs collapses that into one pipeline: a topic goes in, a finished video comes out roughly eight minutes later, with nobody touching anything in between.

Why that's harder than "call some AI models"

Calling an image model, a video model, and a voice model individually is the easy part. The pipeline chains script generation, scene planning, image generation, animation, narration, and sound design into one sequence, and every stage depends on what the last one produced. Making that chain hold together, instead of drifting or breaking silently three stages in, was most of the actual work.

How it's built

Staying visually coherent. The same characters and settings have to survive dozens of separate generation calls without visibly changing halfway through. That meant managing prompts and seeds across the whole run as one continuous state, not per-call. Routing around providers, not to them. Gemini, Replicate (Flux, Seedance, LTX, Minimax), and ElevenLabs each sit behind one interface, so a stage's model can be swapped without touching anything upstream or downstream of it. Knowing the cost before spending it. A pre-flight credit check confirms a run can actually afford to finish before generation starts, instead of finding out four stages in that it can't. Keeping runs from bleeding into each other. Clerk-backed auth and workspace isolation so concurrent runs from different users never share state.

Stack

Next.js · TypeScript · Google Gemini · Replicate · ElevenLabs · FFmpeg

Where it landed

A topic in, a finished video out, about eight minutes later, coordinated across a dozen-plus models with no manual stitching between stages. The days of production work that would normally sit between those two points just don't happen anymore.