AI-driven procedural content generation is not one tool but a family of techniques with very different strengths. The short answer for developers: simulation-based methods suit natural, emergent phenomena like terrain; constraint-driven methods give you firm authorial control over structured content such as levels; machine learning models split by content type — GANs excel at images and 2D levels, while large language models dominate narrative work like quests and dialogue. And before picking any algorithm, you must answer a more basic question: is your content generated offline during development, or online at runtime?
Offline vs. Online PCG: Two Core Use Cases
A survey of PCG in games, including work on emerging LLM integration, draws a line that shapes every downstream architectural choice: generated content can be divided into online and offline generation. Online generation means the content is produced during the runtime of the game itself, while offline generation means it is created during game development. These are not merely two schedules for the same work — they impose different demands.
When content is generated offline, the generator behaves like a developer tool. You can afford slower, heavier computation, and because a human reviews the output before it ships, imperfect results can be filtered by hand. The generator's job is to assist creators, multiplying what a team can produce, not to surprise players on launch day.
Online generation flips the contract. The generator becomes part of the shipped product, and its output must be playable, coherent and fast enough to run inside a live game session. This is where PCG becomes a gameplay mechanic in its own right — the source of dynamic, replayable worlds. Deciding which of these two roles you need comes first, because it constrains which techniques are even viable: a technique that needs minutes of compute or curated training data may work beautifully as an offline assistant and be entirely unusable at runtime.





