This study proposes a method for generating captions that highlight differences between pairs of garment images, with the goal of providing useful information to consumers engaged in product comparisons. To design the content of ideal captions, we analyzed product comparison articles focusing on apparel sold on e-commerce platforms. The analysis revealed that focusing on salient differences between garments and explicitly presenting subjective attributes as recommendation information, based on objective attributes, is an effective approach. The proposed method uses a prompt-based vision-language model (VLM) to estimate attributes step-by-step and generate captions for each garment. A qualitative evaluation of the generated captions, focusing on the degree of difference emphasis and the relevance between attributes, confirmed that the proposed method provides sufficient quality of information to effectively support product comparisons.