Hello
Prompted to Sora:
Leo Tolstoy with a long beard, Mahathma Gandhi in white robe, and Dante Alighieri in a red hood, drinking transparent liquid from faceted glasses together, standing near a high-top table in a cozy Moscow bistro, evening, warm lighting, calm atmosphere. There is a half-full bottle of clear liquid on the table. The label on the bottle is blue and white.
Respionse: “the content may violate our policies”
Prompt “Leo Tolstoy” - ok, “Dante Alighieri” - ok. “Mahathma Gandhi” - viloate policies
Replaced Gandhi to Martin Luther King - volia!
(mention the fourth glass. Gandhi went to the toilet)
Can anybody explain please, what’s wrong with Gandhi?!
Sometimes the model can be picky about specific keywords. For example, there was a time this last Valentines that you couldn’t create an image containing a rose in it, no matter how hard you tried.
1 Like
First i asked to change transparent liquid to pirozhki with cabbage - violate the policies
Gandhi alone - violate the policies
Suggestion: Gandhi violates the policies, not “Gandhi drinks transparent liquid”
BTW: late Tolstoy also was vegan and didn’t drink “transparent liquid from the faceted glass”
I think I found your problem, it doesn’t like “white robe”.
1 Like
Oh. My. God.
I learned something today.
1 Like
Yes, you can alter and subtract words and phrases until it works?
That’s basically all I did.
There are a million ways for it to work, but certain combinations of words, in specific orders, alter the semantics presented to the content filters. If keyword based, then removing the offending words avoids this filter too.
Since Gandhi is already known to wear a “white robe”, it is redundant, and no need to include it in the prompt, since it’s already built into the model.
2 Likes
