(How to) transform a playing field into a great map?

Hi. I have developed a generator that creates a map that defines the terrain of a playing field in 30x20 squares. I think you can guess what the symbols and colors mean. Example:

The map is generated anew each time, so it shouldn’t repeat itself.
I’ll need a minimap for the briefing later on, though the map doesn’t look all that great—it’s rather informal.

I also have a style sheet that is to be used:

The map should look like this, but with the terrain information passed from the generated map.

The AI ​​needs to run locally on a computer and be able to complete the task within 1–2 minutes.
I currently use ComfyDesktop with SDXL and IPAdapter for this:

As you can see, despite the precise task description in the positive prompt, the result is still far from matching the map in terms of both quality and accuracy:

The map needs to be quite precise. Currently, it consists of a 30x20 grid
(though there will likely be more squares later); the properties of these squares, as well as the course of the river, etc., must be clearly and accurately represented.

I still have the map in a sort of text format—basically just as the computer generated it—though I’m not sure if that helps?:

Could someone help me figure out how to implement this?
That would be absolutely fantastic!

Best regards,
Toby

Edit: By the way, this is what ChatGPT conjured up from my map template:

I was busy and didn’t notice it, though I assume it took longer than two minutes—albeit not much longer. In principle, however, it clearly demonstrates that it is possible.

reason one of the map images seems mushy is your running out of memory in late stage detailing. atleast thats how i describe it.
my recomendation - break up map creation in to 2 componets

  1. map architecture
  2. Map Art.

lets those be 2 different cycles. that way you get full generative capacity on the detail stage of creation. see if that helps your workflow.
i love the project though. i have a project for a TTRPG engine, and something like this is a great related concept.

I’ve ultimately decided to generate the map myself using a script. I tried out several AIs, but they either failed to deliver (many struggled with the perspective alone) or produced results of insufficient quality.

I do have another AI up my sleeve for the game, though it serves an audio-related function (no, not music).

As for the solution regarding this thread:
I experimented with masks, height maps, color maps, style templates, and so on. However, the various workflows never yielded the results I had hoped for.
Ultimately, the required output was too specific and rigid for an AI to handle with the necessary quality and high level of accuracy (>99%).

Training a custom, lightweight AI would have been an alternative, but for a map that only exists in the game for about 30 seconds, that level of effort simply isn’t worth it.