Semantic-Aware Intelligent Agent for Multi-Task Image Communications

Abstract

Semantic communication (SemCom) has emerged as a promising paradigm for enhancing communication efficiency in sixth-generation (6G) wireless networks. However, existing SemCom frameworks are often constrained by task-specific architectures, limiting their generalization across complicated scenarios with heterogeneous tasks. To address this limitation, we propose LISA—a large intelligent semantic agent that establishes a unified SemCom framework capable of supporting multiple image-related tasks. LISA comprises two core modules: a semantic module and a channel module. The semantic module integrates a multi-source prompting component, a large language model, a memory component, and an auxiliary toolkit to jointly interpret source images, task requirements, and environmental information. This enables the extraction of task-adaptive semantic representations suitable for transmission, such as visual features for image reconstruction, stylistic descriptors for image style transfer, and class labels for image classification. The channel module employs a vector-quantized variational autoencoder to discretize continuous semantic representations into codewords drawn from a shared codebook at both transmitter and receiver, thereby enabling efficient semantic transmission. The indices of selected codewords are then modulated using a transformer-based modulation scheme, ensuring compatibility with digital communications while enhancing robustness against channel noise. Experimental results demonstrate that the proposed LISA-based SemCom system outperforms state-of-the-art baselines across three popular tasks: image reconstruction, image style transfer, and image classification, while significantly reducing communication overhead. These results highlight the great potential of LISA for practical applications in 6G.