Remastering a classic game means generating new levels that feel hand-authored, and dressing them (as well as the original ones) in modern artwork, without breaking the collision geometry that gameplay depends on. This talk presents a complete, production-tested pipeline built for the commercial remaster inspired by a 1993 platformer: (1) structural level generation, systematically comparing character-based LLM workflows, vector-based LLM approaches, and diffusion models; (2) autonomous validation through semantic touchup and a BDI playtesting agent in a fuzzy physics world; and (3) a multi-layer visual finalization module using fine-tuned LoRA and image-manipulation models that realizes a novel target art style while strictly preserving geometry. A user study (N=65, 1,300 ratings) shows diffusion-based generation matching human-authored difficulty at a fraction of the authoring time, and finalized artwork that players could not distinguish from the original human-made art, where off-the-shelf image models consistently break game colliders. Beyond results, the talk shares what it takes to move generative AI output from “impressive demo” to near-shippable game content.
1. The advantages and disadvantages of level representation (character grid vs. vector description vs. image), and concrete evidence that diffusion beats token-based LLMs for spatial level structure, at drastically lower generation time.
2. A practical architecture for autonomous playtesting: a heuristic BDI agent that validates generated levels for playability, difficulty and other metrics, without human testers.
3. Why off-the-shelf image models break game geometry, versus our geometry-preserving multi-layer finalization recipe (conditioning + fine-tuned LoRA restyling + layer compositing) that produces artwork players can’t tell from human-made.
4. How to design a user study that meaningfully evaluates generative game content (1,300 ratings across aesthetics, difficulty, human-likeness).
5. The overarching lesson: decomposing generation into dedicated, strictly constrained modules is what turns generative AI output into near-shippable commercial content.
