I’m observing different unwanted behaviors when using the DALLE-3 API. Today, the API generated an image containing parts of the literal revised prompt:
“Generate an intricate landscape embodying…”
This is beyond unwanted keywords being shown in the generated image. The image contained parts of the literal prompt.
Its being literally creative maybe
Jokes apart, I am certain it is a bug. For now try mentioning in your prompt the action it must not take.
Long term, I am certain this will get tweaked to work the opposite way, prompt should be included only if asked.
I’ve been noticing this too. Will report to the DALLE team, although I’m sure they know and are working on it. Thanks for pointing it out, though.
I’m encountering this quite frequently - maybe 1/10 images. Most recently with multiple lines of text. Thanks for your quick reply, please let me know about any updates.
It also misspells most labels and sometimes pots labels on the image even when instructed not to include them.