Day 09: AI animation mistakes I caught in one day, from a boot off the foot to a founder saying the narrator's line

Some mistakes take a trained eye. Others take a person who has worn shoes. On Day 09 I mostly needed the second kind of person.

Day 09 at a glance

What kind of AI animation mistakes does a director catch? On my film, on October 12, mostly the physical and the obvious: a boot being laced in midair, a founder lip-syncing a line the narrator says, clips I had already judged shown to me again. Claude rendered 73 clips on RunPod for about $9.17, I approved 17 of the 52 I judged, and we rebuilt the score and two scenes in Premiere around my notes.

What went wrong overnight?

Midjourney stopped us before breakfast. Since the block on Day 02, Claude makes frames through a paced runner, one job at a time. That runner had made 111 jobs in 24 hours, 58 of them between midnight and 9am, and Midjourney asked the account to complete a CAPTCHA to show it wasn't a robot. The runner stopped itself. Only I'm allowed to solve a CAPTCHA.

I didn't believe it. After the fake CAPTCHA on Day 07, I asked it to double-check before I believed it.

This one was real, and the fix was pacing. The runner now does nothing from 1:00 to 8:30 in the morning, no more than 4 jobs an hour, and a daily cap that I set at 42 and raised to 60 that evening for speed. I let it run past its limit once while I was at the desk, and I was clear about that:

"don't remove the limit, just this is an exception because i am here"

Later I put it plainly. A slower system that acts more like a human is how we keep Midjourney happy with us.

Why does an AI animate a boot that isn't on a foot?

Scene 02 has John getting ready for work, the same moves every morning. The shooting script said he laces his boot. It never said the boot was on his foot. So the frame showed a boot on the floor, and the animation had John tie the laces of an empty boot and move on to his lunch tin.

The second clip is the founder of the Megaplex, Augustus Blackwell, holding up oversized scissors to cut a ribbon on archive tape. "Some cities have a heart. We have a founder." That's the narrator's line. The prompt carried the line, so the founder said it with his own mouth. I wrote on the clip: "just remove the text line from the prompt."

My note in chat was simple: the shot was supposed to be John putting the shoe on.

Both fixes are now in code. The script for that shot says the boot is already on John's foot. The render tool only attaches a line and its audio when the speaker is in the shot and isn't the narrator. And before anything is animated, the motion prompt is read against the actual frame, so every action starts from what the frame shows.

Why did I change Lucy back?

On Day 07 I asked for a sexier Lucy, and Claude redesigned her. Two days later I watched the clips and saw the problem:

"we have too many good clips of the old lucy to be using the new lucy who is sexier, we need to resolve that"

So Lucy went back to her original locked design. The newer face is retired, and so is every frame made with it: 29 frames across Scenes 02, 03, 07, 09 and 21, approved or not. I told Claude the old approvals didn't count after a character change. The clips I had approved from the old Lucy all stay.

It sounds like a step backward. It's the cheapest fix of the week. A character who changes faces between scenes costs more than a character who is a little less glamorous than I imagined.

Why was I judging the same clips twice?

Around lunch I asked Claude why it kept showing me, and re-rendering, shots I had already approved.

The phone review page was stale. Its list of clips hadn't been rebuilt since October 10, so 113 clips that were already judged or covered came back to me, including more takes of shots I had already solved. The rule now is that the page is rebuilt from the Mac review before it ever opens, and once a shot has an approved clip, its other takes disappear.

An hour later I said what I had been thinking all week: the film was still in 30-second and 1-minute pieces, we have about 60 scenes, and we were still on the first ten. That had to change.

What changed in the edit?

Overnight I asked for two sound fixes. The stems Claude used had long silences in them, and some scenes had dead air with no music at all. A new tool now trims silent stretches by whole beats, so the pulse keeps going, and fills gaps with the previous cue at a lower level. Scene 08, the Penthouse, had about 31 seconds with no score. Now it has about one.

Claude also built a "Blade Runner grade" with deep blacks and more contrast. I said no:

"i want the haze ... blade runner was more thematic not color"

It was reverted within the hour. The soft orange and red haze is the look of this film. Blade Runner is the grit, the rain and the violence, not a color setting.

Then Scene 08. The spy was hoisted, then standing, then hoisted again, and the cut skipped his death. My notes: "you missed the part where he dies. we wanted the gruesome as much as possible." The rebuild puts the violence in order. He talks on his feet first, there is one yank, he hangs in every shot after it, and his death plays to the end of the approved clip.

After I watched the scenes in Premiere I had three more notes: a missing voice, approved clips that weren't in the cut, and music that started and stopped "as if the thought was lost." The missing voice was the cowboy at the poker table. His line, "Cheatin' tin can!", had never been recorded. It's in now, and the music was rebuilt so each scene carries fewer, longer musical thoughts across its cuts.

When Claude handed me a list of four edit decisions for Scene 11, I sent them back: "do what you think is best for the story based on what Orson Welles would do." That's the job now. It decides, logs why, and I judge the result.

The one I liked

Mike in Scene 05, laughing as he seals the deal with John for that night: "Deal! After the whistle." I approved it on the phone. Earlier in the week the AI kept drawing Mike as the wrong man until I insisted on his name and his picture. Here he is, the right man, in the cut.

The numbers for October 12

• RunPod: 10 batches, about $9.17, 73 clips. The $0.28 batch launched empty because of a bug that was fixed the same day.
• Clip verdicts: 52 judged, 17 approved, about 33%. Across the whole film so far: 143 of 507 approved.
• Frames retired with the old Lucy design: 29.
• Lip sync, measured on the clean source: of 93 lines, 10 could be read by the meter, and 8 of those are in sync.

So what did Day 09 teach me?

An AI crew can render 73 clips in an afternoon and still lace a boot in midair. Most of my notes on October 12 weren't about taste. They were about cause and effect: feet go in boots, narrators don't have faces, dead men don't stand back up. Each of those notes became a rule in code, which is the only way I've found to stop the same mistake twice.

The other lesson is about me. I changed Lucy, then changed her back. I asked for Blade Runner and then said I meant the story, not the color. The AI follows my words very literally, so my words need to be better. Follow the run on the Making Death Match series page.

FAQ

What AI animation mistakes did the director catch on Day 09?
A boot laced while it sat on the floor, a founder lip-syncing the narrator's line, an out-of-order hoist in Scene 08, and clips I had already judged shown to me again.

Why do AI video characters say the narrator's lines?
The animation model lip-syncs whatever line and audio the prompt carries. The fix was to attach a line only when its speaker is on screen and isn't the narrator.

How much did one day of AI video rendering cost?
About $9.17 on RunPod on October 12 for 10 batches and 73 clips, including one empty $0.28 batch caused by a bug.

Why was Lucy changed back to her old design?
There were too many good approved clips of the old Lucy. The newer design was retired with its 29 frames, so she looks the same in every scene.

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. It's more forgiving than I was about that boot. Download Halfsies on the App Store