A Reinforcement Learning and Transfer Learning Synergy Framework for Scene Generation in Clothing E-Commerce Visualization

Abstract

Virtual display of clothing is a key link to improve consumers' online shopping experience, but traditional solutions rely on manual design, which is inefficient and difficult to meet individual needs. This study proposes an innovative reinforcement learning-transfer learning collaborative generation framework (RL-TL Synergy Framework) for intelligently generating diverse and high-quality clothing virtual display scenes. Our methodology employs an Actor-Critic reinforcement learning architecture. This is synergized with a transfer learning component that leverages pre-trained convolutional networks (e.g., ResNet-50) on large-scale aesthetic datasets (e.g., AVA) and our curated dataset of historical successful displays (FashionDisplay-20K, containing 20,000 expert-annotated scenes) to extract and transfer prior knowledge on layout and style. The core of the framework lies in the deep integration of the sequential decision optimization capabilities of reinforcement learning (RL) and the knowledge reuse capabilities of transfer learning (TL): reinforcement learning agents learn the optimal clothing matching, spatial layout and perspective rendering by interacting with the environment. strategy to maximize visual appeal and user conversion rate; The transfer learning component is responsible for extracting transferable prior knowledge such as layout aesthetics and color coordination from historical display data or related scenes (such as home furnishings and art displays), and efficiently transferring and adapting to new tasks, significantly reducing the exploration cost of reinforcement learning and sample requirements. Experimental evaluation on held-out test sets demonstrates that our framework improves scene generation efficiency by approximately 40% compared to a pure reinforcement learning baseline. It outperforms rule-based and Generative Adversarial Network (GAN) baselines by a significant margin, achieving a Structural Similarity Index (SSIM) ≥ 0.91, an increase in aesthetic score of 18.7%, and a measured improvement in simulated click-through rate of 23.8% in an A/B testing environment. The framework effectively accommodates variations across store aesthetics, seasonal themes, and user preferences, thereby delivering an efficient, intelligent, and scalable virtual display generation solution for fashion e-commerce platforms.