|
Abstract:
|
This work investigates the application of Euler-based Conditional Flow Matching (CFM) to the
procedural generation of tile-based 2D game maps, framed as a Constraint Satisfaction Problem
(CSP) in which generated maps must satisfy implicit local and global structural rules learned
from a training corpus. A generative pipeline is implemented from first principles in C# with
TorchSharp: a two-level U-Net parameterizes a time-dependent vector field trained via the
Conditional Flow Matching objective, and inference proceeds through an Euler ordinary
differential equation solver at fifty function evaluations per sample. The Video Game Level
Corpus Super Mario Bros distribution provides fifteen levels across three biomes (Overworld,
Underground, Treetop) as the training set. Development proceeded across four iterations, each
targeting the dominant failure mode observed in the previous iteration's output as measured by a
six-category failure-mode analyzer. Iteration 1 established the baseline pipeline; Iteration 2
explored class-balanced loss variants and biome conditioning; Iteration 3 tested
capacity-and-duration scaling and produced a negative result identifying a fundamental quality
ceiling of the CFM plus argmax decoding formulation; Iteration 4 introduced a deterministic
post-generation repair pipeline of eleven rule-based primitives. Across thirty generated chunks,
the hybrid pipeline reduced three hundred and nine unintended structural violations to zero, with
sixty-five intentional Treetop ground gaps preserved as a biome feature. |
