Measuring Product Quality Using Images: The CLIP Q-Score and an Application to Real Estate

This paper introduces the CLIP Q-score, a novel method for extracting objective quality metrics from visual data using Contrastive Language-Image Pre-training (CLIP). The authors leverage the ability of multimodal neural networks to map images and text into a shared embedding space. By comparing an image against three specific text prompts—representing luxurious, ordinary, and dilapidated states—the model calculates a polarity score that reflects the perceived quality of the product.

The researchers applied this technique to approximately 500,000 images from a major Russian real estate platform. They validated the CLIP Q-score by demonstrating its high correlation with assessments from multimodal large language models and its alignment with institutional knowledge, such as the geographic distribution of property quality in Moscow and the impact of state-sponsored demolition programs on housing scores.

The study demonstrates that the CLIP Q-score is a robust predictor of real estate market outcomes. Properties with higher image-based quality scores command significantly higher prices, even after controlling for traditional hedonic characteristics like square footage, location, and building age. Furthermore, the authors find that higher scores are associated with increased market liquidity, meaning these properties spend less time on the market before being sold or rented.

Traditional hedonic pricing models often struggle to quantify subjective visual attributes like interior design or state of repair. The CLIP Q-score provides a scalable, open-source, and privacy-preserving tool for researchers to incorporate visual information into economic analysis without the need for expensive manual labeling or proprietary AI inference services.