Out There door E, "not the guy", made on base MiniMax H3 on our rented Vast boxes, 2026-11-08

This ad is 78 seconds long, made from 7 takes of one man in his parked car, every word from the locked door E script with his own generated voice.

The prompts came from our own writer, which copies the way fal writes prompts for H3 and asks for one continuous shot per take.

What the checks found

• All 7 takes speak 96 to 100% of their lines, and the only differences are spellings such as "mid-50s" and "who will".
• The same man, blue t-shirt, black cap and car appear in every take, and the Out There stick reads clearly on the product takes.
• The framing jumps between takes: take 4 is an extreme close-up while the others sit at arm's length, and the QA's colour check read that as a change of shirt.
• Take 5 has two camera cuts inside it, the man pauses about a second between some sentences, and his hands are malformed in places.

The full QA verdict is here: 6 fatal, 30 major, 31 minor, and four of the fatal lines are the framing change described above.

One frame from each take beside the reference frame

Render: 4 rented RTX PRO 6000 boxes, 29 minutes from rent to the last take, each box destroyed when it ran out of work.

Every version of the same take, 2026-11-08

All of these speak the same locked line and come from the same reference frame of the same man. Numbers 1, 2 and 3 were each generated separately, so each drew its own composition and gestures. Numbers 4 and 5 start from the exact file in number 1, so they differ from it only by what the upscaling step did.

Number 4 is the floor, because plain stretching invents nothing. Number 3 is what costs 2.85 times as much to generate.

Numbers 5 and 6 both came back at 960 wide, which is larger than the native 768 in number 3, so they will look bigger on screen whether or not they recovered real detail. Judge them on the beard and the hand.