Modality-Aware Unequal Error Protection for Multimodal Generative Semantic Communication

Abstract

Efficient transmission of multimodal semantic features under strict physical-layer constraints is a key requirement for generative semantic communication. However, existing approaches often treat different semantic features uniformly despite their unequal impact on perceptual quality, which leads to inefficient resource allocation and semantic collapse. To address this issue, we propose Modality-Aware Unequal Error Protection (MA-UEP), a framework for multimodal generative semantic communication over wireless channels. To our knowledge, this is among the first works to explicitly formulate modality-aware physical-layer resource allocation under realistic fading conditions. Different modalities play fundamentally different roles in generative reconstruction: text governs global semantics, while structural modalities such as edge and depth provide noiserobust spatial guidance. Based on this insight, MA-UEP jointly optimizes modality subset selection and power allocation under a strict total transmit-power budget. Without additional model training, MA-UEP significantly improves semantic consistency and perceptual quality, particularly at low signal-to-noise ratio (SNR), by adapting the modality subset and power split to the channel state. Simulation results on the Cityscapes and COCO datasets demonstrate that MA-UEP outperforms conventional uniform allocation under the same candidate modalities and identical power constraints. On Cityscapes, it raises the CLIP (Contrastive Language-Image Pre-training) score at an SNR of -2 dB from 0.581 to 0.827. These results establish modalityaware resource allocation as a practical design principle for generative semantic communication.