產品背景生成的質量評估

Abstract

本文介紹了一個名為 QAPBG (Quality Assessment for Product Background Generation) 的評估框架，專門用於評估產品背景生成結果的質量。雖然這個領域在工 業界有廣泛應用，但卻缺乏結構化的評估方法。此框架提供了一個寶貴的工具， 用於全面評估產品背景生成領域中的生成模型，其特點是包含六個不同的指標: 背景提示對齊、產品一致性、產品孔洞與背景一致性、合理的地面產品、自然陰 影和美學質量。每個指標對於提高生成圖像的真實感和效果至關重要，既考慮了 視覺吸引力又考慮了功能一致性。我們提出並嚴格評估了評分這些指標的方法， 從而建立了一種標準化的方法來評估和改進產品背景生成。

This paper introduces an evaluation framework, named QAPBG (Quality Assessment for Product Background Generation), tailored specifically for assessing the quality of outcomes in product background generation—a domain with widespread industry applications yet lacking structured evaluation methods. This framework serves as a valuable tool for the comprehensive evaluation of generative models in the product background generation domain, featuring six distinct metrics: background prompt alignment, product consistency, product hole and background consistency, reasonable grounding product, natural shadow, and aesthetic quality. Each metric is crucial for enhancing the realism and effectiveness of generated images, addressing both visual appeal and functional coherence. We propose and rigorously evaluate methods for scoring these metrics, thereby establishing a standardized approach to assess and improve product background generation.