TL;DR

  • On October 21 the render queue tried to start a batch 95 times. A safety check refused every one, so no GPU was rented and no clip was made. It cost $0, and a day.
  • The check exists to keep fight scenes fast. It refuses any action shot whose prompt also says "softly" or "slowly". Two prompts tripped it. One had "a slash of orange light" across Elektra's eyes. The other had a projector "throwing" its beam down a table.
  • The edit room tried a dissolve from one take of John into another. For a moment the second John's whole head sat inside the first John's nose. A critic caught it and the dissolve came out.
  • I approved 36 of 46 frames in 11 minutes. Claude later held three of them for fixes, including a padlock seal that says MEGAPLEX in wax.
  • Claude's pre-judge threw out 103 of 232 new frames, including four takes where a gang member called a Viper became a toy Dodge Viper.

Day 17 was about how much one rented GPU can make in an hour. Day 18 was about the day it never got the chance.

Day 18 at a glance

Claude runs the crew on my film. A runner sends prompts to Midjourney, one at a time. A render queue sends the frames I approve to rented GPUs on RunPod to be animated. Before each batch, a preflight check reads every prompt. The edit room re-cuts one scene every hour or two, and critics score each cut. Claude's pre-judge looks at every new frame before I do.

All times here are Eastern. The Mac runs on Bangkok time, so I convert.

Owner-approved clip, animated later: Elektra with a slash of orange light across her eyes, the prompt the preflight read as a knife fight
Owner-approved clip, animated later: Elektra with a slash of orange light across her eyes, the prompt the preflight read as a knife fight

How did two words stop the render queue?

On October 17 some fight clips came back slow and soft. Part of the reason was in Claude's own prompts. It had used calming words to stop the picture from drifting. So we added a rule to the preflight check: if a prompt describes action and also uses a calming word like "softly" or "slowly", the whole batch is refused. Write fights fast, beat by beat.

On October 20 I approved two frames. One is Elektra in close-up with a hard slash of orange light across her eyes. The other is Dr. Waters at the end of a briefing table, his hand on a slide projector. Neither one is a fight.

Claude wrote a description of each frame for the animation prompt. Elektra "speaks softly" and the camera "pushes in very slowly." The projector beam is "throwing a cone of light," and the camera "pushes slowly." The check saw "slash" and "throwing" as action, then saw "softly" and "slowly," and refused.

It refused every batch, not just those two shots. Each batch carried up to 16 approved frames. The queue retried every few minutes with the same prompts. On October 21 that happened 95 times. It started at 6:53 p.m. on the 20th and ended at 7:08 a.m. on the 23rd: 334 refusals in all.

The check did its job. It kept the money in my pocket, $0 spent. What it didn't do was tell anyone. The health check only said "no new clip for 4.6 hours." One edit room run saw that a new batch had started and wrote that the queue was working on it. That batch was refused too.

On the 23rd the edit room found the cause. It rewrote the two descriptions, and seven more like them. The first pod started at 7:16 a.m. I approved both clips at 11:24 and 11:25 a.m. that day. The health check now names any description that would trip the preflight.

Owner-approved clip, animated on October 23: Dr. Waters at the slide projector, the beam the check read as a throw
Owner-approved clip, animated on October 23: Dr. Waters at the slide projector, the beam the check read as a throw

How did John end up inside his own nose?

Scene 04 opens on John jogging down the stairwell of his building, then cuts to him on the packed train to work. Both takes are ones I approved. They were made on different days, from different frames, so they're two slightly different drawings of John.

On its 4:30 p.m. run, the edit room tried a dissolve between them instead of a hard cut. In the first take John's face fills the frame on its last half second: his nose, mouth and brow. The second take has him smaller, chest up, among the workers.

Halfway through the dissolve, at 4.79 seconds, the small John's whole head sits inside the big John's nose. The big John's eyes are above it. His mouth runs across the small John's chest.

The round-two critic caught it, frame by frame. I had already ruled out a dissolve like that between two other takes of John. Two designs of the same face morph into each other. The edit room went back to a hard cut, with the train sound starting half a second early to carry you over it.

Scene 04 went from 5.83 to 6.17. The rest of that run moved lines onto the faces that say them, and gave Dot's toy drone its hum as it whips past John's ear.

Claude's edit, dropped: at 4.79 seconds of the dissolve, one John's head inside the other John's nose
Claude's edit, dropped: at 4.79 seconds of the dissolve, one John's head inside the other John's nose

What did I judge on October 21?

One sitting, from 11:11 to 11:22 a.m.: 46 frames for Scenes 21 and 24 to 27, most of them the Kyoto hotel job. I approved 36 and rejected 10. Three of the rejections were love-scene frames, which stay off this diary. Late that night I also approved six clips from the love scenes and the fighters' party, made on the Mac's own GPU.

Before an approved frame is animated, Claude writes a plain description of what's in it. That afternoon it described 34 new ones, and it held three of them for a fix:

Each one gets repainted, then comes back to me.

Owner-approved frame, held by Claude for a repaint: John with the padlock, "Megaplex" pressed into the wax seal
Owner-approved frame, held by Claude for a repaint: John with the padlock, "Megaplex" pressed into the wax seal
The prop lock for the Megaplex mark: a circle crossed by three lines, no lettering
The prop lock for the Megaplex mark: a circle crossed by three lines, no lettering

What did Midjourney get wrong?

The runner finished 59 jobs, from 12:07 a.m. to 11:58 p.m., with its usual night pause from 2:06 to 9:59 p.m. Claude's pre-judge checked 232 frames for Scenes 25 to 28 and passed 129. The 103 it failed never reached me. Some of them:

Many of the rest were John with black hair. His lock says light sandy buzz cut.

REJECTED by Claude's pre-judge: a toy Dodge Viper on the table where a Viper gang member should be
REJECTED by Claude's pre-judge: a toy Dodge Viper on the table where a Viper gang member should be

What did the edit room do in ten runs?

Ten runs started on October 21. Six stood down because the run before was still going. The other four each took one scene. The numbers are the critics' average out of 10, before and after.

The numbers for October 21

So what did Day 18 teach me?

A rule that only says no is half a rule. The preflight was right not to send those prompts to a GPU. It was wrong to stay quiet about it for two and a half days while everything looked fine.

The words were the same lesson from the other side. "Slash" and "throwing" were about light. A check that reads words without the sentence will find fights everywhere. Follow the run on the Making Death Match series page.

FAQ

Why did the render queue rent no GPU on October 21?
A preflight check refuses any batch where an action shot's prompt also uses a calming word. Two prompts said "softly" and "slowly", and words about light, "slash" and "throwing", made them look like action. Every batch was refused, 95 times that day. Nothing was spent.

Why not dissolve between two takes of the same character?
In AI animation, two takes are two drawings. In a dissolve they overlap, and on October 21 one John's head showed up inside the other John's nose. A hard cut hides the difference between them.

Why hold a frame the director already approved?
I approve for style and story. Claude checks details before animation: readable text, which turns to mush in motion, and whether each character matches their lock. Three of 36 frames went back for a repaint.

What is Claude's pre-judge?
It looks at every new Midjourney frame before I do and throws out the clear failures. On October 21 it failed 103 of 232, including four toy cars.

While you wait for the next update: play Halfsies

Halfsies is our iPhone game, live now. One cut. Two halves. Each day you get three objects and one attempt at each: cut it exactly in half by eye and the scale tells you how close you got. No restart needed. Download Halfsies on the App Store