Encoding “Liubai”: An Aesthetic Perception Framework and Differentiable Metrics for Chinese Ink Wash Style Textile Pattern Generation

Abstract

This paper presents a computational framework for quantifying aesthetics of Chinese ink wash and applying them to generative models. We define differentiable metrics for the three core elements: the compositional balance of “Liubai” (negative space), the calligraphic quality of “Bichu” (brushstroke), and the tonal diffusion of “Moyun” (ink wash). Using these metrics, we benchmark unpaired image-to-image systems—CycleGAN, MUNIT, ChipGAN, and diffusion pipelines with controllable methods (Style LoRA, ControlNet-Tile, IP-Adapter)—on photo-to-ink transfer. Results show a trade-off: diffusion excels at “Moyun” texture fidelity, while ChipGAN with explicit aesthetic losses better preserves “Liubai” and “Bichu” structure. The study also highlights limitations of generic image-quality metrics (e.g., FID) for artistic evaluation. We further validate implications for phygital textile design via seamless-tiling tests and small-scale physical samples. Finally, we outline a unified, material-aware scheme embedding fabric diffusion physics (Fick’s law) into a Physics-Informed GAN objective to jointly optimize aesthetic fidelity and printability.