Application of multimodal deep learning in art design element extraction and advertising visual optimization

In today’s digital media environment, advertising increasingly relies on AI to create visuals that are both aesthetically appealing and commercially effective. This research introduces a multimodal deep learning framework that combines image, text, and behavioral data to improve advertisement design. The research uses a Kaggle open-source dataset of 1000 samples, with simulated behavioral metrics limiting real-world generalization. Using features extracted through ResNet50 and Word2Vec, the proposed Galactic Swarm Optimized–Bidirectional Encoder Cycle-Consistent Generative Adversarial Network (GSO-BE-CycleGAN) model enhances color harmony, layout balance, and artistic style while preserving the original marketing message. GSO-BE-CycleGAN achieves superior visual quality, engagement accuracy, and training stability over existing models. A dataset of 1000 artwork and advertisement images with text descriptions was used to train and evaluate the system. Experimental results demonstrate improvements in visual quality (PSNR 30.14 dB, SSIM 0.94) and engagement-related performance using proxy simulated indicators (CTR 0.731, F1-score 0.825), outperforming multimodal transformer baselines under controlled experimental settings. It is important to note that CTR, interaction rate (IR), and conversion-related metrics were synthetically generated based on predefined behavioral distributions and therefore serve as proxy simulated indicators for comparative evaluation. These results demonstrate methodological robustness and multimodal optimization capability rather than direct validation of real-world commercial effectiveness. Future work will involve validation using large-scale real advertising campaign datasets.