I tried this but it's almost not possible, I think because the training data of these AI models doesn't contain many pictures with this perspective.

the result is always bad.