These are just observations I've made; perhaps other people will have other observations/solutions.
I plan to continue adding to the article If anyone has any other ideas to share, I'm all ears.
Architecture and urban planning can pose many problems, among which I particularly note the scale of the characters, aerial view (when you don't want), arch, pillar (especially when there are buildings at the corners), domes, arcades, squares, but also gardens (especially with Illustrious models), not to mention the numerous detail problems.
It occurred to me that some problems might also stem from the images ratio.
First, I like to find whimsical or extravagant architectures and I find that the SDXL models are generally easier to use.
For the character scale, the only way I know is to make loras including characters, which is not always easy.
Example of "correct" proportion between people and buildings (But with a specific character, it's sometimes difficult to achieve):
For the aerial view, the only solution is obviously to play with the negative prompts, the same for pillar and arch, The problem may stem from the images used to train the LoRa of the model used, or from a combination of both (as below).
Note also the number of domes.
Domes pose several problems, the first that comes to mind is "finding oneself" under a dome, meaning indoors, the only solution I know is to add (outdoor) to the prompts, pillar and arch to the negative prompts (I find they often go together); using dome one can also find small domes in various places, garden in particular, for which I have no solution.
Below three images with multiple glass domes, "glass domes" in the prompts.
The problem with arcades is that, for example, there may be gardens where there should be buildings.
Sometimes when generating urban ensembles I've managed to get squares or things resembling them, but it is always by chance.
Two examples:
As for the gardens, that only applies to the Illustrious models, indeed, using "garden" gives a good probability of having a garden of potted plants, and this is even more obvious when adding "plant" and "flower", the solution I found "garden" and "grass", but in my experience, other prompts can play a role for exemple describe a character has added flowerpots!
In the two images below, the only difference in the prompts is "grass" in the second image, in the first image, we'll notice the problem (building) with the arcades on the right.
In town, I have only rarely and unintentionally creating good square and I'm not sure I'll succeed even with (if anyone has an idea, I'm all ears).
Furthermore, I'm not sure if AI can differentiate a column, a pillar, or a monolith, or even to know what machicolations or a watchtower are.
In the interiors there can be asymmetry, among other things with columns and vaults, not to mention scale.