An AI Video Production Pipeline with a Six-Gate Automated QC Harness

A studio producing a weekly episodic series with two recurring characters wanted to stop paying for cloud generation and run the whole production on their own RTX 5090 server. The interesting part was never the generation; it was consistency. Across shots, across episodes, the same character has to remain the same character, and a human watching every frame is not a scalable quality process.

The pipeline

We worked over SSH on the client's own machine and built a Python pipeline that takes a scripted scene from text to a finished 1080p cut:

• Keyframes generated from the client's own character LoRAs.
• Motion generated image-to-video from those keyframes.
• Voice — cloned first, then replaced with purpose-built English canon voices for both characters, locked so later episodes stay on-model.
• Lip-sync applied per character with masks, so two speakers in one frame do not contaminate each other.
• Assembly to a 1080p master with ffmpeg.

We also fixed the client's pre-existing pipeline scripts, reviewed the server environment, and wrote up the workflow and model choices so the studio could operate it without us.

The QC harness — the part worth talking about

Generative video fails in specific, repeatable ways: a mouth that keeps moving when the character is listening, a face that drifts into a different person between shots, a style mismatch between cuts, a blur patch where a mask edge sat. So rather than eyeballing renders, we wrote a gate harness that scores every cut automatically before a human ever sees it — mouth-motion gating, style matching against a reference, identity and face-lock checks, voice quality analysis, and an overall video QC pass, rolled up into a six-gate verdict.

That harness is what turns "it looks off" into a numbered, reproducible defect. Every review pass in the project is recorded as numbered findings against the gates, which is why the defect list is long — it is supposed to be.

Status: ongoing

A milestone encode was produced for client review, and an M2 master subsequently passed all six gates. Review then logged three further defects — identity splitting across shots, listener lip motion, and a visible blur patch — and those remain open. The engagement is live and the scope has been restructured into measured milestones against a written pass/fail metrics grid, which is the right shape for work of this kind: acceptance is defined numerically up front rather than negotiated after the render.

We are not presenting this as finished work. The rendered output is the client's intellectual property and is not published here.