Conditional GAN-Based System for Automated Packaging Design and Market Demand Alignment Using Multi-modal Evaluation Networks

Abstract

This paper introduces a cGANs-founded system designed for the autonomous creation of packaging designs to meet real-world market demand. The system architecture involves three different levels: a data acquisition level that aggregates multimodal data from brand design banks, consumer feedback, Enterprise Resource Planning (ERP) and Customer Relationship Management (CRM) systems, as well as trends from social media platforms; a cGANs-based generation level that uses a U-Net generator with a PatchGAN discriminator to create design products that are sensitive to market trends and heritage brand identity; and an evaluation level that uses a dual-stream network with ResNet-50 to extract image features and BERT to analyze semantic user feedback. The dataset used for this work comprised 10,000 labeled packaging images with 500,000+ entries describing both structural and unstructured consumer behaviors. The experimental results achieved best performance compared with baseline approaches, achieving a Fréchet Inception Distance (FID) of 28.4 and an Inception Score (IS) of 12.7, exhibiting a high quality and diversity of auto-generated images. Furthermore, the model achieved a 9.1/10 brand consistency score, 15% improvement in consumer satisfaction, and 20.2% increased conversion rate of purchases with A/B tests. The results prove that there is validity to the argument that this proposed system can create high-quality and brand-consistent packaging designs to suit market demand.