T2V, an autonomous video production pipeline.

Constraint

Producing a single educational video required four to five roles for a small studio, and one to two weeks of human effort per video. Off-the-shelf tools could write a script. Off-the-shelf tools could generate images. None of them was a pipeline. Each tool solved a slice; nothing tied them together into a system that ran the work without a human between every step.

First principles

A video is a sequence of decisions made under attention constraints. Which topic. Which hook. Which line of the script. Which frame for which beat. Which thumbnail. Each decision is small and reversible. The problem is not that any one decision is hard. The problem is that there are hundreds of them, and a human in the loop at each one collapses the throughput to linear-in-headcount.

If the decisions can be made automatically against a fixed quality bar, and the failures are cheap to retry, the throughput uncouples from headcount. Cheap retries are the unlock.

Pipeline

The operator triggers a new production from Telegram. The pipeline runs to completion and returns the cut for review.

Frame-chaining

The experimental motion track runs Seedance 2.0 plus GPT Image 2 with explicit frame-chaining. Chunk N's end frame becomes chunk N+1's start frame. This is what avoids the identity drift that killed earlier face-based attempts. Gemini 3 Pro runs as the video judge against the rendered output.

Economics

$13 per finished video at current draw. The cost-down work targets $6 by replacing two paid steps with in-house alternatives. The numbers matter because they decide what is worth doing at all.

Where it came from

T2V was built for the lab's own marketing surface. The first autonomous video shipped 2026·03·24 at 8:49 running time. The channel has been running continuously since then. The case study on the channel is documented under deployments.