<p>The deployment of deep learning-based visual communication systems in cultural creative scenarios demands both high perceptual quality and real-time responsiveness, yet edge hardware imposes strict constraints on computation, memory, and energy. This paper proposes a unified framework that jointly optimizes attention mechanism design and resource consumption for edge-deployed visual communication in the cultural and creative industries. First, a Lightweight Dual-Branch Attention (LDBA) module is introduced, which decouples channel recalibration and spatial modeling into parallel branches using depth-wise separable convolutions and grouped strip pooling, achieving over 565 × compression in floating-point operations relative to standard self-attention. Second, a dynamic resource-aware scheduling strategy based on a greedy heuristic is developed to assign tasks to edge nodes according to a composite urgency score that integrates task priority, deadline constraints, and real-time node utilization. Third, a device-edge collaborative inference framework with adaptive split-point selection is designed to partition model execution between terminal devices and edge servers in response to fluctuating network conditions. Experiments on the ArtBench-10 and a self-constructed Cultural Creative Product (CCP-5K) dataset demonstrate that LDBA achieves accuracy within 0.3 percentage points of full Transformer attention at roughly one-third the inference latency, while the scheduling strategy reduces latency by 35.9% and energy consumption by 33.7% compared with static allocation. The collaborative inference framework maintains robust performance across a wide bandwidth range, validating the practical viability of the proposed approach for real-time intelligent visual services in cultural creative venues.</p>