Frame 1
The tools that turn a weekly hook list into talking-actor variants are desktop and phone apps and a per-second avatar API, with nothing in between that a Linux CI runner can call, so the pipeline ends at a Mac mini under a desk.
Frame 2
Each variant names a script, a licensed actor, a voice and the brand kit. The runner posts it nightly for most teams, and each variant returns as a captioned 1080 by 1920 MP4 billed in finished seconds.
Frame 3
The endpoint captioned, cut and exported footage studios already had. The generation half came in spring 2026, when open-weight audio-driven actor models became good enough to fine-tune on licensed footage and serve ourselves.
Frame 4
A text-to-speech model voices the script, and an audio-driven video model fine-tuned only on licensed performer footage generates the actor frame by frame. A 30-second variant is about 900 frames, roughly two GPU-minutes on today's model.
Frame 5
A take below 0.80 on audio-visual sync is re-rendered twice at most, then returned failed and unbilled. A script flagged above 0.85 for claims ad platforms reject waits for a named person to release or drop it. Each decision is a labeled record for that workspace's screen, and it trains the screen all customers use only if its owner opts in.
Frame 6
Forty variants is 80 GPU-minutes inside an hour, then close to nothing until the next night. Manifests carry next week's hooks and unreleased campaign footage that studios keep from model vendors, so our models run on cloud GPU capacity we control in Oregon and no copy is kept outside the United States.
Frame 7
It runs on cloud GPU capacity we control in Oregon. By March 2027 a distilled version replaces it, because the per-second price holds only if two GPU-minutes per 30-second variant comes down. By September 2027 an embedding-retrieval index holds scripts close to ones a workspace already dropped before they render.
Ran growth at a mobile-app studio where the nightly variant job lived on a Mac mini under his desk; when it died over a long weekend, the winning ad ran a week past fatigue with nothing to replace it.
Spent five years running render farms for animation studios and could not see why a talking-actor ad needed a desktop machine when every frame of it comes out of a GPU somewhere else.
Reconciled paid-social media plans against creative invoices at a DTC brand and never once got a credit balance to add up to the seconds on the plan.