Earticle

This paper proposes a prompt–optimization framework for generating style–consistent game images using Stable Diffusion XL. Given a reference game item image, the system first extracts an initial prompt using a vision–language captioner and a domain–specific prompt bank. The extracted prompt is converted into a list of noun-like elements, and a genetic algorithm searches for compact combinations of these elements under dual gating based on SSIM and CLIP scores. The best combinations are treated as a “style template” that can reproduce the reference image with high structural and semantic similarity. We then investigate whether this template can be reused when the main object is changed while preserving the original visual style. Experiments on fantasy-style item images show that the framework reconstructs reference images using only 8–9 automatically discovered prompt elements, and that changing the main object token together with associated detail elements yields image sets that share a consistent visual style. In contrast, naïvely replacing only the main object token often produces visually ambiguous or stylistically inconsistent images. These results demonstrate that combining automatic prompt extraction from images with evolutionary optimization provides a concrete example of style–preserving prompt design for game item image generation.