Beyond Isolated Products: Vision-Language Reasoning for Aesthetic Product Recommendation

Abstract

In retail environments, effective product visualization is paramount for driving sales and customer engagement, necessitating advanced methods for showcasing inventory. A significant challenge in current digital commerce is the cognitive disconnect between isolated product images and a customer's unique space. In such scenarios, shoppers struggle to mentally project how isolated items will look together, which leads to aesthetic uncertainty, purchase hesitation, and high return rates. To bridge this gap, we present AestheticVLR, an end-to-end framework for aesthetic product recommendation that integrates vision–language reasoning with structured semantic decomposition and constraint-aware synthesis. Given a user-specified theme, the framework combines Large Language Models (LLMs) for intent and attribute reasoning with Vision-Language Models (VLMs) for visual compatibility assessment. This process enables coherent recommendations across multiple product categories by using generated reference scenes that act as semantic anchors to guide multi-modal synthesis. The final output presents users with a curated list, offering at least three to four optimally suited inventory options for every necessary product type. By automating the alignment of disparate inventory items within a unified aesthetic context, AestheticVLR enables seamless theme-to-purchase conversion and significantly enhances the digital shopping experience through data-driven, aesthetically grounded decision-making, while the exact name of the underlying generative model are withheld for business reasons.