Years after initial Early Access release, Generative AI has finally reached the point where this game can be taken to the next level. The technological improvements were on two fronts: Text generation, and image generation.
Text generation
When LLMs first came out, they could barely code. Nowadays we expect them to write fully compiling code every time and if they so much as misinterpret some ambiguous word we said we get mad at them :)
What this means for AI Roguelite 2D: I created a scaffolding, a JSON schema somewhat approaching a “coding language” for the LLM to generate, as well as all the game logic to parse and execute said JSON structure. The LLM can now be instructed to output entire legitimate JSON structures of worlds including items, crafting recipes, entity abilities, how those abilities work, what status effects do, etc. This makes for much more emergent gameplay than the previous iterations. For example:
- “Knockback” isn’t an effect and isn’t even mentioned, but the AI successfully generated it by designating how to move the target entity after it’s hit by a specific ability.
- “AOE calm” isn’t a specific effect, but the AI generated it by specifying the affected area and a behavior override that temporarily prevents entities from attacking.
Image generation
We went from models that turn words into images based on vague signals from training data, to models that appear to actually understand your query and are able to follow specific, detailed instructions on what elements to put where.
What this means for AI Roguelite 2D: I discovered that an advanced model can output up to 20 animations per image while keeping track of what each row is supposed to be. That is an incredible leap forward from regular image generation, and it’s also the only way I was able to keep animation costs low enough to make available in any reasonable quantity. Note: AI-generated animations are opt-in, and you must enable them in Options.
Caveats
- Animations and world generations have strict daily limits due to the high cost. These limits may be relaxed in the future.
- Neither text generation nor animation generation has perfect accuracy, and it goes without saying that even when working correctly, animations won't look nearly as good as real human-made one.
- Initial generated worlds by default have 3-5 biomes and cannot yet be expanded (option to expand existing worlds is a planned feature)
- This is a massive overhaul so there will be many bugs which will need to be fixed over time.
- Generated worlds are highly unlikely to be fully balanced because they’re generated by an LLM with no testing (however, most times they are not completely broken, and the LLM seems to have common sense about what’s reasonable).
- The text gen model sometimes takes a long time to generate the whole world.
Other major changes
- The game is now focused on power creeping by exploring, gathering, and crafting, not by leveling up.
- Terrain generation is now procedural instead of squares
- The game is now in 2.5D (billboarded sprites in 3D) rather than 2D
- Overhauled inventory UI and HUD
- Entities may belong to a faction. Factions are enemies of each other and will often fight each other. Each faction is assigned an enemy/neutral/ally relationship with the player.
Planned features
- Generating more biomes within an existing world
- Editing a world by prompting the AI (e.g. “Please make the first biome less difficult” or “Please add an AOE fire to sparks attack”)
- Building static objects
- More complex interactions
- Tentative: Reinstate on-demand LLM calls, towns, and/or talking NPCs, which had been removed because they didn’t add much and were imbalanced (only if I find a way to make it fun)