Edge Caching‐Enabled Artistic Painting Style Transfer

ABSTRACT The proliferation of mobile applications and augmented reality (AR) has surged the demand for real‐time artistic style transfer. However, existing neural style transfer (NST) algorithms predominantly rely on heavy convolutional encoders to extract style features from reference images during every inference cycle. This paradigm presents a critical bottleneck for mobile edge computing (MEC): it incurs high latency due to repetitive computation and consumes significant bandwidth by requiring the transmission of high‐resolution style images for every request. To address these challenges, this paper proposes a novel framework: edge caching‐enabled vector‐quantized disentangled style transfer (VQD‐ST). Unlike traditional methods that map artistic styles to a continuous and infinite feature space, our approach learns a discrete, compact Style Codebook in the cloud. This codebook encapsulates a rich diversity of artistic prototypes through a vector quantization (VQ) mechanism, allowing complex style features to be cached directly on edge nodes. We further introduce a cached fixpoint disentanglement strategy to ensure that the cached style prototypes remain strictly independent of user content, preventing semantic leakage. Experimental results on the MS‐COCO and WikiArt datasets demonstrate that VQD‐ST achieves a stylization speed of 32.5 FPS on edge devices (NVIDIA Jetson Nano) and reduces bandwidth consumption by over 99% compared to state‐of‐the‐art arbitrary style transfer methods. By shifting the computational burden of style extraction to an offline cloud phase and enabling style retrieval at the edge, our method establishes a new state‐of‐the‐art trade‐off between visual fidelity, inference latency, and network efficiency.