DALL-E3 GPT inpainting: Generation harder than revision

Hi all, when I use the in-painting function to revise my generated image, I discovered that “generate something new in a certain area” often fail compared to “replace colors or erase something”. For instance, when asking
Dall-E to “generate a new hat for the rabbit”, 10 out of 10 times it fails. When asking it to “replace the rainbow with a house” then 2 out of 3 attempts it erased the rainbow but not adding anything and the third time gives a cloud-like house that is not satisfactory.

So far the highest success chance I have would be “replacing eye colors, dress of a person etc”. Anyone could help to understand why?