is it better to have many captions or fewer detailed captions?

[Xibanya]

When creating a dataset for training a DALLE model, if I have a photo of a city skyline at night, is it better to have in the text file

• a photo of a city
• a photo of a skyline
• a photo at night

or

• a photo of a city skyline at night

or

• a photo of a city
• a photo of a skyline
• a photo at night
• a photo of a city at night
• a photo of a city skyline at night
• a photo of a city skyline at night
  (etc)

And if option 3 is the best, if I want training to go faster and don't want to go the route of listing every single permutation on the phrase, which of the first two options is better?