Edge‐Collaborative Inference Mechanism for Animation Style Transfer in Resource‐Constrained Environments

ABSTRACT The rapid proliferation of mobile Augmented Reality (AR) applications has necessitated real‐time, high‐quality image processing to merge virtual animation elements with physical environments seamlessly. However, performing sophisticated animation style transfer tasks that require significant computational power to maintain structural integrity and stylistic consistency remains a challenge for resource‐constrained mobile devices. This article proposes AniEdge‐Net, an innovative deep adaptive device‐edge collaborative inference network designed to optimize animation generative tasks. Building upon the principles of partitioning neural networks, AniEdge‐Net introduces a Semantic‐Preserving Partitioning (SPP) module to balance computational latency and semantic entropy. To minimize transmission overhead, a Dual‐Path Content Encoding (DPCE) module is utilized, implementing 1‐bit quantization and structural re‐parameterization to achieve a 123× feature compression ratio. Furthermore, a Bandwidth‐Aware Style Configuration (BASC) module dynamically adjusts transmission parameters to accommodate varying network conditions. Finally, a Temporal–Spatial Consistency Compensation (TSCC) module on the edge server utilizes adaptive instance normalization to recover stylistic details, bridging the information gap caused by extreme compression. Experimental results demonstrate that AniEdge‐Net achieves a total latency of approximately 61.8 ms with a marginal accuracy loss of less than 3%, outperforming conventional edge‐only and device‐only inference paradigms in animation‐specific scenarios.