A solo developer’s second iteration on 12 months of work solving the hard problem in AI game art: game-ready assets that share one consistent style across an entire world. The title undersells the mechanism — the clarification in the comments is that the fix is a LoRA-finetuned image model, so every generation inherits the same look.
The workflow, in practice:
- A LoRA-trained model anchors the style; vague prompts like “a spooky forest village” produce strong results without detailed prompting.
- Shadows and lighting detail are baked into the generated worlds — which is why stock engines can’t render them properly.
- The approach ships with its own engine, open-sourced under MIT (capybara_2d_engine).
- It works standalone or end-to-end: build the world and assets, hand the folder to a coding agent for gameplay, and sync changes back and forth.
- Resulting games load fast and distribute across web, mobile, and desktop.
The honest part of the writeup is the architecture: consistency came from fine-tuning, and the real engineering problem turned out to be a renderer that can display baked-in lighting. AI pipelines still end at engine problems.
The workflow is also explicitly built around agent handoff — the coding agent is the gameplay programmer while the human curates the art direction. A useful reference point for anyone hitting the style-drift wall in AI-assisted asset pipelines.