I'm making a diffusion model mapgen.
It's MOSTLY done with agentic code, so it's SLOP CODED just a tiny bit less than a 100%.
I've managed to make it a study project for my university course.
I've already managed to generate some slop like this:

Help me by annotating maps: https://faf-mapgen-atlas.elven-bowl-7506.chatgpt.site/
Github: https://github.com/LimeZ3/Lime-Map-Generator

Some additional slop:
MapGen: an open-source study project exploring neural networks for FAF map generation
Hi everyone,
I’m working on MapGen, a study project for my AI Automation course at Fontys in the Netherlands, exploring whether neural networks can learn useful map structure from existing Supreme Commander: Forged Alliance maps.
I intend to keep the project open source and clearly labelled as a study project, with its code, experiments, progress, and limitations available for the community to follow.
The main question is simple: can a model learn something about what makes a map interesting to play, beyond generating terrain that looks plausible?
Perlin noise is useful for generating terrain variation. A complete procedural map generator adds rules that shape terrain, place resources, and create playable layouts. The possibility I want to investigate with neural networks is learning some of those relationships from existing maps: how terrain, starting positions, expansion opportunities, reclaim, and access routes fit together.
For example, could a model eventually help generate a map with contested expansions, several viable attack routes, or a particular balance between land and naval gameplay?
That is a research question, not a proven advantage. Neural networks could also learn bad patterns, repeat familiar layouts, or produce terrain that needs substantial correction. Playability checks and actual games would still be necessary. This project is not a claim that AI already produces better maps than existing generators such as Neroxis.
There is already work behind the idea: the project includes tools for extracting terrain and gameplay data from existing maps, preparing it for machine learning, running baseline experiments, and collecting community annotations. It is still experimental, and I’m not presenting it as a finished map generator.
This is where community experience can help. Extracted data can describe terrain and resource positions, but it does not automatically explain how a map feels to play. Does it encourage raiding? Does it favour defensive play? How much does naval control matter?
I’m preparing a map annotation page where players can contribute those judgments. The intention is to use them to help organise the dataset and investigate whether player descriptions can guide generation. Different opinions are useful too; a map can play very differently depending on the players and settings.
On the personal side, I have 10+ years of gaming experience, and I care about this game and the project beyond the course assignment. My intention is to keep playing and developing this after the course ends. Development will have to fit around study and life, but I want to make that commitment visible through public progress updates, documented next steps, and an open repository.
I’d appreciate feedback from players, map makers, and generator developers—especially on this question:
What makes a FAF map worth playing again, and what should a generator learn to get right?