One day, all nine clips I generated were broken.
Not a poor hit rate. Every single one.
A moon turned into a human face. Screens filled with rectangular blocks. A full day of generation, gone.
My first theory was wrong
Here was my initial explanation:
“Dark images with little detail must be fragile when an AI animates them.”
It sounds reasonable. Starfields, night seas — I had indeed been using dark source images. So I decided to pick brighter material.
That theory was later completely disproved.
I measured the actual darkness of 18 source images.
Successful ones averaged 62.6% dark pixels. Failed ones, 62.0%.
Essentially identical.
The clincher was a snowfield image: 0% dark, bright to the point of clipping.
It failed too.
Had I trusted my hunch, I would still be selecting material on the wrong criterion.
Catching two culprits separately
So I stopped guessing and ran a controlled experiment.
Two variables, four clips.
- Prompt: with “do not draw ~” / with zero negative wording
- Seed (the number setting the randomness): the one from the disaster / a different one
The result was strikingly clear.
| disaster seed | different seed | |
|---|---|---|
| with “do not draw” | moon becomes a human face + rubble | rectangular frames across the screen |
| no negatives | moon intact, minor colour blocks | completely clean |
Two readings:
1. The rectangular glitch came from the prompt.
Hold the seed, change only the wording, and it appears or vanishes.
2. The moon-into-a-face came from the seed.
Hold the wording, change only the seed, and it appears or vanishes.
They were two independent problems. And on the day everything died, I had stepped on both at once.
Write “do not draw it” and it gets drawn
This was the biggest lesson.
Trying to prevent breakage, I had written:
Do not produce rectangular blocks or glitch artifacts
But the model I was using had the mechanism for receiving negative instructions switched off.
So what survives? The words “rectangular blocks” — read as an instruction to draw them.
The line I wrote to prevent damage was causing it.
I changed how I write:
- no negative wording
- describe what is in frame, starting from the subject
- instead of “do not shatter,” write “remains one smooth unbroken circle“
One more thing
The same investigation taught me something about motion.
Anything that “appears,” “detaches,” or “falls” will break.
Snow falling, a leaf leaving a branch — these cannot be handled in a short generation budget. Motion where objects come into or out of existence is hard for the model.
I now keep motion within “the surface flows” and “the shape changes slowly.”
Silent failure is the scary part
The most frightening thing here was elsewhere.
The broken clips passed as “successful” in the system.
Only completion was being checked, so a ruined interior produced no error.
I considered detecting it by file size, but the numbers for broken clips sat within the normal range. Undetectable that way.
Which means there was no method other than looking.
The horizon showed up again
One more finding.
Source images with a horizon line broke more often.
One horizon composition was tried three times under varied conditions and failed all three.
I had previously learned, with still images, that a horizon is a chair for man-made objects. It turned out to be a weak point in video too.
The same weakness surfacing somewhere else. That happens.
Losing a day of generation hurt.
But by resisting the urge to act on a hunch, and instead narrowing the experiment to two variables, I finally reached the real cause.
Had I stayed with “dark images break,” I would still be choosing source material by the wrong rule.
Worse than the failure itself is deciding the cause by instinct. That bill comes later.
This article does not guarantee any medical or health benefit and is not a substitute for professional medical care. If you have any health concerns, please consult a physician or other qualified professional.
