Ask for a picture of space, and a torii gate appears.
It sounds like a joke. It cost me most of six months.
This is not about sound. It is about the visuals that accompany the tracks. But in production time, it took just as much as the audio did.
Things I never asked for keep arriving
Healing tracks need backdrops. Starfields, seas of cloud, deep space. I ask an AI to paint them.
And the results contain things I did not request.
- torii gates
- city skylines
- spacecraft
- for some reason, a giant hand
- human figures
I specify “quiet cosmos,” and there is a small town sitting on the horizon. At thumbnail size you do not notice. I have shipped images like that.
Writing “do not include” summons it instead
The obvious first move was negative wording.
“Do not draw a torii.” “No buildings.”
It made things worse.
Depending on how the generation works, negatives can be inactive — and merely naming the thing calls it into the picture.
The time I wrote “no rectangular blocks” and got back an image full of neat rectangular blocks, I had to laugh.
What 34 images told me
I stopped going on instinct and laid the failures out and counted.
Reviewing 34 images, the pattern was unmistakable.
All 13 contaminated images had a horizon line.
The cause was not the word “torii.” It was composition.
Give the model a horizon and it wants to place something on the boundary between ground and sky. What gets summoned to sit there is a gate, or a building.
A horizon is a chair for man-made objects.
Once I understood that, the fix changed. Not more negatives — remove the horizon itself.
Drop the ground-level viewpoint for an orbital one. Or look straight down at water.
Contamination dropped sharply.
Even so, never skip the check
It happened again recently, on a seasonal article’s cover image.
I asked for a single leaf touching a pond at dusk. The first attempt put in a pavilion and a human figure; the second, a distant tower.
The third, with the horizon removed and the camera looking straight down, finally came out clean.
Even when the generator reports success, look at it yourself. That rule still stands.
There was a batch where 8 of 15 assets were contaminated. More than half. Skipping the check would have published all of them.
The same thing happens with sound
The lesson applies to audio too.
Once you automate generation and processing, “done” reports stop being trustworthy. The numbers can pass while the actual result falls apart on listening.
So the final step is always the same:
listen with my own ears / look with my own eyes.
Automate what can be automated. But do not automate the judgement.
That one came from failure.
More of my time goes into checking that nothing unrequested slipped in than into making beautiful images or sounds.
It is mundane work. Skip it and it always comes back 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.
