ArtStroke-GAN: A Unified Framework for Interpreting Watercolor Aesthetics via Personalized Stroke-Based Graphic Language

Painting is an expressive art form traditionally associated with considerable technical skill and prolonged training, making it largely inaccessible to the general public. Computational stroke-based rendering (SBR) approaches have emerged to simulate human-like artistic processes digitally; however, current methods encounter notable limitations, such as spatial inconsistencies, stylistic discontinuities, and insufficient control over fine-grained stroke details. To address these challenges, this paper introduces ArtStroke-GAN, a novel generative adversarial network framework designed to synthesize realistic and personalized watercolor paintings. The proposed framework employs a modular three-part GAN architecture consisting of a stroke-generation network, a dedicated color-enhancement module, and an adversarial discriminator. Adaptive spline-based stroke modeling and iterative, semantic-aware stroke optimization strategies are integrated, enabling progressive refinement from coarse to detailed artistic representations. Experimental evaluations conducted on diverse image datasets (CelebA and ImageNet) demonstrate that ArtStroke-GAN achieves superior performance compared to state-of-the-art methods, including Stroke-GAN, Paint Transformer, and Neural-Paint, across quantitative metrics such as LPIPS, FID, style loss, and content fidelity. Qualitative assessments further confirm its ability to produce aesthetically coherent, structurally consistent, and stylistically diverse watercolor renderings, significantly advancing the capability of computational art generation frameworks.