Exactly. The whole "hobby vs real work" framing is just a fancy way of saying you're paying by the hour to train a model that you don't own, on a problem it wasn't built to solve. Your "Orion Sync" morning cost your company something, even if it was fun.
Making a template isn't giving up. It's just stopping the meter from running on a taxi going the wrong way.
Your stack is too complicated.
Right, because the model's weights aren't going to change just because you put your internal jargon in quotes. It's like expecting a dog to understand French because you said the command with an accent.
The "generate background, overlay text" method works, but the real value isn't just consistency, it's version control. When "Project Blue Sky" inevitably gets renamed to "Project Azure Horizon" in six months, you just edit the text layer in your template. No re-training the team on a new set of fragile prompt incantations.
You're right about it being a manual override every time, and that's the real kicker. For a one-off graphic, fine. But if you're documenting an entire platform with dozens of these terms, writing a unique rule for each one is unsustainable. It becomes its own maintenance headache.
Yeah, welcome to the club 😅 I ran into the same thing trying to make a badge for our "Greenwave Deployment" dashboard. Got nothing but ocean scenes.
It's a known thing, like others said. The sandwich trick or misspelling might help a tiny bit, but it's super hit-or-miss.
Honestly? For a wiki graphic, I'd just grab a nice, simple abstract background from the AI and then add the text in something like Canva. Saves so much headache. What kind of graphic are you trying to make?
Quotes won't help. The model's training data has millions of "blue sky" images and maybe three internal codenames. It will always default to the highest probability token, which is the literal sky.
You've hit a core limitation. Treating the model like a graphic design intern who just doesn't "get it" is the wrong mental model. It's a pattern matcher with zero context about your company.
For wiki graphics, the reliable fix is to separate the tasks. Generate a neutral, abstract tech background with the AI. Then add your proper names as text in a basic image editor. It's less clever, but it works every time and you can update the text in five seconds when the marketing team changes the name again.
shift left or go home