Papers
Topics
Authors
Recent
Search
2000 character limit reached

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

Published 2 Aug 2026 in econ.EM and cs.CV | (2608.01544v1)

Abstract: The CLIP Q-score is a novel, safe, fully reproducible, and computationally efficient method for extracting objective product quality metrics from visual data using contrastive language-image pre-training. We introduce the technique and provide an extensive application to real estate data from an online platform (∼500,000\sim500,000 images). Our open-source metric aligns with LLM assessments and proves to be a powerful predictor of housing market prices for both sales and rentals. We also show that a higher CLIP Q-store is associated with better liquidity (reduced time on the market), especially for properties on sale.

Summary

  • The paper introduces the CLIP Q-score, a transparent image-based measure of product quality, that uses a pretrained CLIP model to convert visual information into a scalar quality index.
  • The study applies this method to the Moscow real-estate market, finding that the Q-score is associated with listing prices and transaction speed for both rentals and sales.
  • Higher Q-scores are linked to higher rental prices (5.8% increase per 0.1 unit) and sale asking prices (3.5% increase per 0.1 unit), indicating that visual quality influences market valuation.

Research objective and contribution

The paper introduces the CLIP Q-score, a reproducible image-based measure of product quality, and evaluates it in the Moscow residential real-estate market. The central methodological claim is that a pretrained CLIP model can convert qualitative visual information into a scalar quality index without task-specific model training, manual image annotation, or external inference services. The empirical application uses approximately 500,000 listing images from the Russian platform Cian.ru and examines whether visual quality is associated with observable housing characteristics, asking prices, and time on the market (2608.01544).

The contribution is therefore both methodological and empirical. Methodologically, the paper proposes a transparent scoring protocol based on contrastive image–text similarities rather than a black-box LLM judgment. Empirically, it shows that the resulting measure is associated with conventional indicators of housing quality and contributes substantial predictive information in hedonic models for both rentals and secondary-market sales. The paper also makes the stronger claim that visual quality affects market liquidity, although the estimated relationship depends materially on conditioning on the initial asking price.

Construction of the CLIP Q-score

CLIP consists of separate text and image encoders whose outputs are projected into a common embedding space. Training uses image–text pairs and a symmetric contrastive cross-entropy objective: matched image–caption pairs receive high similarity, while mismatched pairs receive lower similarity. At inference time, the model can therefore compare an image with alternative textual descriptions and assign relative similarity scores without requiring a newly trained quality classifier.

The proposed Q-score uses three deliberately ordered prompts:

  • “This is a photo of a luxurious apartment.”
  • “This is a photo of an ordinary apartment.”
  • “This is a photo of a dilapidated apartment.”

For each image, the image embedding is compared with the embeddings of the positive, neutral, and negative prompts. Similarities are multiplied by CLIP’s learned temperature parameter and transformed through a three-way softmax. The final score is the polarity difference between the positive and negative probabilities:

Qi=Pi(+)−Pi(−).Q_i = P_i(+) - P_i(-).

The neutral prompt affects the score indirectly by entering the softmax denominator. This construction produces an interpretable bounded measure: higher values indicate greater relative alignment with the positive description, while lower values indicate stronger alignment with the negative description. The method can be generalized by adding further textual categories, although the scalar polarity measure necessarily compresses potentially multidimensional concepts such as luxury, maintenance, architectural style, and photographic presentation.

The implementation uses the open-source openai/clip-vit-base-patch32 model. Images are resized and centrally cropped to 224×224224 \times 224 pixels, with 32-pixel ViT patches. This model choice prioritizes computational efficiency over visual resolution. The preprocessing decision is consequential: centrally cropping can remove peripheral features, and the relatively coarse ViT configuration may be less sensitive to fine-grained interior details. The authors address this concern empirically by residualizing the score against technical image characteristics.

Data and measurement design

The data consist of Moscow listings created between July 1 and August 14, 2025, followed until December 5. The sample contains 16,047 rental advertisements and 11,129 secondary-sale advertisements with images. The corresponding image totals are 275,477 for rentals and 216,271 for sales. Listings include unit-level characteristics such as area, rooms, bathrooms, ceiling height, floor, and reported repair condition; building-level characteristics such as construction year, materials, and demolition-program status; and geographic variables measuring distances to the city center, public transit, green areas, and industrial zones.

Listings without images are excluded, comprising approximately 7% of rental advertisements and 5.5% of sales advertisements. This restriction is defensible for an image-based analysis, but it changes the estimand: the results concern properties represented visually on the platform rather than all listed properties. Moreover, the data contain asking prices rather than completed transaction prices. The paper argues that this may make estimated price associations conservative if visually attractive properties confer greater bargaining power, but that interpretation cannot be tested directly without transaction data.

The authors calculate several conventional image-quality indicators—resolution, aspect ratio, Laplacian-variance sharpness, BRISQUE, and mean HSL brightness. All are statistically associated with CLIP polarity in large samples, but their explanatory power is limited: the regression R2R^2 is only 0.025 for rentals and 0.030 for sales. Thus, more than 97% of the variation in raw CLIP polarity remains unexplained by these technical image attributes. The paper interprets this result as evidence that the score is not reducible to blur, brightness, resolution, or cropping-related artifacts.

For the main analysis, the authors use an adjusted score defined as the property-level average residual from this image-quality regression. This adjustment improves the separation between semantic content and technical presentation, but it does not eliminate other forms of confounding. For example, staging, camera angle, selective image choice, renovation style, and the number and ordering of images may all influence the score while also reflecting seller strategy or unobserved property quality.

Validation against external and institutional benchmarks

The first validation exercise compares CLIP scores with ratings produced by the multimodal LLaMA4 model. The LLM was prompted to assign a 0–10 luxury score to each photograph. Scores were obtained for approximately 90% of advertisements, corresponding to 14,129 rental and 10,226 sales listings. Processing required approximately two weeks and cost about $250. The distributions are similar and the cross-method relationship is strongly positive, supporting the claim that CLIP captures a semantic notion of visual quality rather than only low-level image statistics.

The comparison also establishes an important operational distinction. LLaMA4 scoring is relatively easy to specify but is slower, costly, dependent on an external provider, and not guaranteed to be exactly reproducible even at temperature zero. CLIP inference can be performed locally with fixed parameters and deterministic preprocessing. Consequently, the paper’s reproducibility claim applies more convincingly to the proposed scoring pipeline than to API-based multimodal assessment. It nevertheless depends on preserving the exact model checkpoint, tokenizer, prompts, preprocessing, and aggregation procedure.

The descriptive evidence is consistent with the intended interpretation of the score. Larger units and apartments with multiple bathrooms receive higher scores. Higher floors, taller ceilings, and more sophisticated reported renovations are also positively associated with the measure. Properties in buildings scheduled for demolition have markedly lower scores. Geographic patterns correspond to the established socioeconomic structure of Moscow: central, western, and northwestern districts score higher, while eastern and southern districts score lower.

The association with construction year is particularly consistent with the historical context. Surviving pre-war properties tend to have undergone substantial refurbishment; much of the post-war Soviet housing stock is assessed less favorably; and post-transition construction receives higher scores. The resulting U-shaped pattern is plausible, but it remains a descriptive validation rather than an identification strategy. Construction year is correlated with location, building type, renovation selection, and socioeconomic composition, so the pattern cannot by itself establish that CLIP measures intrinsic housing condition.

Predictive content in hedonic models

The principal empirical test embeds the adjusted average Q-score in log-linear hedonic regressions containing extensive unit, building, geographic, and district fixed-effect controls. The coefficient is positive and highly statistically significant in both markets:

Market Q-score coefficient Interpretation of a 0.1 increase
Long-term rentals 0.582 Approximately 5.8% higher rent
Secondary sales 0.350 Approximately 3.5% higher asking price

The magnitudes are economically substantial. A 0.1 increase in the score is described as roughly corresponding to a 10% higher model-assigned probability of the high-quality description relative to the low-quality description. Under the paper’s interpretation, visual quality is not merely correlated with listing appeal; it captures valuation-relevant information omitted from conventional structured fields.

Including the Q-score increases adjusted test-sample R2R^2 from 0.874 to 0.886 for rentals and from 0.897 to 0.904 for sales. Test RMSE falls from 0.179 to 0.170 for rentals and from 0.192 to 0.186 for sales. These improvements are modest in percentage-point terms but meaningful given the already rich feature set and the high baseline predictive performance. They support the narrower claim that image-derived semantic information provides incremental predictive value conditional on standard hedonic characteristics.

The findings are robust to replacing OLS with histogram-based gradient boosting. Test adjusted R2R^2 rises from 0.887 to 0.899 for rentals and from 0.915 to 0.923 for sales when CLIP is added. RMSE declines from 0.173 to 0.164 for rentals and from 0.181 to 0.172 for sales. The similar incremental contribution across linear and nonlinear specifications indicates that the result is not an artifact of an inappropriate linear functional form, although the partial-dependence analysis suggests mild nonlinearities in the score–price relationship.

Feature-attribution analysis places the Q-score among the most important predictors. Area and distance to the center dominate; the Q-score ranks third for rentals and approximately fourth or fifth for sales, depending on the predictive model. This ranking is notable because the comparison set includes detailed structural and geographic information. It also qualifies the paper’s strongest interpretation: visual quality is important, but it is not the primary determinant of housing prices in this dataset.

Liquidity and time on the market

The paper extends the analysis beyond prices by examining listing duration. Closed rental listings remain online for a mean of 27.5 days, compared with 51.3 days for closed sales listings. Right-censoring is substantially higher for sales—18.1% versus 3.8% for rentals—because the observation window ends before more sales listings are removed. The authors use Kaplan–Meier methods and proportional-hazard models to account for censoring.

Unconditionally, higher-Q-score properties appear to remain listed longer. This seemingly contradictory result is explained by the positive association between visual quality and asking price: higher-quality properties occupy more expensive market segments, which may transact more slowly. Once the initial asking price and other characteristics are included, the estimated hazard ratio for CLIP exceeds one. Conditional on price, therefore, higher-scoring properties are removed from the platform more quickly, particularly in the sales market.

The implication is that image quality has two opposing observed relationships with liquidity. It is associated with higher prices, and higher prices reduce the unconditional rate of listing exit; holding price fixed, however, image quality improves the conditional hazard of sale. The paper reports that this effect is economically significant for sales but not for rentals. This distinction suggests that visual quality may matter more for buyer search and evaluation in durable-asset transactions than for rental decisions, although the mechanism is not identified.

The interpretation of “time on the market” requires caution. An advertisement’s removal need not imply that the property was sold or rented; it may have been withdrawn, expired, replaced, or temporarily delisted. The authors mitigate this issue by revisiting closed advertisements for an additional 14 days to detect reopening, but the outcome remains a platform-based duration measure rather than verified transaction time.

Limitations and open questions

The paper’s principal results are predictive and associational. The positive Q-score coefficients do not establish that improving a property’s visual quality would causally increase its price. Unobserved renovations, furnishings, staging, seller professionalism, property maintenance, and selective image disclosure may affect both the score and market outcomes. District fixed effects and extensive controls reduce, but do not resolve, these concerns.

Prompt dependence is another substantive limitation. The Q-score is not an intrinsic property of an image alone; it depends on the semantic contrast between “luxurious,” “ordinary,” and “dilapidated,” as well as on language, prompt syntax, model checkpoint, and preprocessing. The paper states that any sufficiently contrasting descriptors should work, but does not provide systematic sensitivity analyses across prompt families, languages, CLIP architectures, or alternative positive–negative anchors. This leaves open whether the reported magnitudes are stable under reasonable measurement choices.

The central crop and base ViT configuration may also discard relevant visual information. Interior images can contain quality cues near image borders, and floor plans or exterior photographs may not be equally informative under the same prompt set. The residual adjustment addresses technical image quality only partially; it does not test robustness to alternative crops, higher-resolution models, image ordering, or aggregation rules.

Finally, the external LLM validation is incomplete as a criterion-validity test. Positive correlation with LLaMA4 demonstrates convergent algorithmic measurement, not agreement with human expert assessments or observed construction quality. The LLM sample also excludes approximately 10% of advertisements and uses at most ten randomly selected images for longer listings. A remaining question is whether the Q-score predicts verified transaction prices and durations after accounting for seller selection, image selection, and independently assessed physical condition.

Conclusion

The paper presents a compact and operationally attractive procedure for extracting a reproducible semantic quality measure from images. In the Moscow real-estate application, the CLIP Q-score aligns with multimodal LLM ratings, institutional housing patterns, neighborhood differences, demolition status, and reported repair characteristics. It also adds predictive information to rich hedonic models: a 0.1 increase is associated with approximately 5.8% higher rents and 3.5% higher sale asking prices, while improving out-of-sample prediction in both linear and gradient-boosting models.

The evidence supports the conclusion that CLIP-based image representations contain economically relevant information not captured by standard listing attributes. The stronger claims—whether the score measures physical quality independently of presentation, and whether visual quality causally changes prices or transaction speed—remain open because the study lacks transaction-level outcomes, human ground truth, and an exogenous source of variation in image-derived quality.

Paper to Video (Beta)

No one has generated a video about this paper yet.

Sign Up to Generate All Videos Subscribe on YouTube

Whiteboard

Explain it Like I'm 14

1. What is this paper about?

This paper introduces a new way to measure the quality of a product by looking at pictures of it. The researchers call their measure the CLIP Q-score.

They demonstrate the method using almost 500,000 photographs of apartments in Moscow. The main idea is that computer software can study apartment pictures and estimate whether a home looks luxurious, ordinary, or run-down.

The researchers then ask whether this picture-based quality score is connected to real-world results, such as:

  • How expensive the apartment is.
  • Whether it is easier to rent or sell.
  • How long its advertisement stays online.

2. Main research questions

The paper tries to answer several important questions:

  1. Can a computer measure product quality from images?
  2. Does the CLIP Q-score agree with human or AI judgments about quality?
  3. Does the score reflect the actual condition of an apartment, rather than just the quality of the photograph?
  4. Do apartments with higher scores have higher rental and sale prices?
  5. Does apartment quality affect how quickly a property is rented or sold?

In simple terms, the researchers want to know whether a computer can look at an apartment photo and notice information that buyers, renters, and real estate agents also care about.

3. How did the researchers do the study?

How CLIP works

The method uses a computer model called CLIP. CLIP was trained using millions of pictures from the internet together with their written descriptions.

It has two main parts:

  • One part studies images.
  • The other part studies text.

Both parts turn what they see into lists of numbers called embeddings. These lists act like coordinates on a map of meaning. Pictures and descriptions with similar meanings are placed close together on this map.

For example, a photograph of a football player holding a trophy should be close to the words “a football player holding a trophy,” but far away from an unrelated description such as “a picture of a forest.”

This training process is called contrastive learning. A simple analogy is a matching game: the computer is shown many pictures and sentences and learns to match the correct pairs while separating incorrect pairs.

Creating the CLIP Q-score

For every apartment photograph, the researchers gave CLIP three possible descriptions:

  • Positive: “This is a photo of a luxurious apartment.”
  • Neutral: “This is a photo of an ordinary apartment.”
  • Negative: “This is a photo of a dilapidated apartment.”

CLIP calculates how well the photograph matches each description. It then turns those matches into probabilities.

The final score is:

CLIP Q-score=probability of positive description−probability of negative description\text{CLIP Q-score} = \text{probability of positive description} - \text{probability of negative description}

A high score means the image looks more luxurious. A low or negative score means it looks more run-down.

The researchers averaged the scores from all the pictures belonging to the same apartment.

Studying Moscow apartments

The researchers collected advertisements from the Russian real estate website cian.ru. The data covered apartments advertised in Moscow between July 1 and August 14, 2025.

They collected information such as:

  • Apartment size.
  • Number of rooms and bathrooms.
  • Floor number.
  • Ceiling height.
  • Type of renovation.
  • Building age and materials.
  • Distance from the city center and metro stations.
  • Asking price.
  • How long the advertisement remained online.
  • All available photographs.

They kept following the advertisements until December 5, 2025, so they could estimate how long each property stayed on the market.

Checking whether the score is trustworthy

The researchers used several checks:

  • They compared the CLIP score with scores produced by another AI system, LLaMA4.
  • They compared it with information supplied by advertisers about the apartment’s condition.
  • They checked whether neighborhoods known to be expensive or prestigious received higher scores.
  • They tested whether buildings scheduled for demolition received lower scores.
  • They controlled for ordinary features such as size, location, number of rooms, and building age.

They used statistical prediction models to see whether the CLIP score improved predictions of apartment prices. One model was a traditional regression, which estimates how different factors are related to price. They also used gradient boosting, a machine-learning method that combines many small decision trees, like asking many simple experts and combining their answers.

The researchers also used Shapley values. These estimate how much each feature contributes to making the model’s predictions accurate. It is similar to working out how much each player contributed to a team’s success.

4. Main findings

The score appears to measure real quality

The CLIP Q-score was strongly related to other signs of apartment quality.

Apartments received higher scores when they had features commonly connected with better housing, such as:

  • More space.
  • More bathrooms.
  • Higher ceilings.
  • Higher floors.
  • More expensive or sophisticated renovations.

Apartments in buildings scheduled for demolition had noticeably lower scores. Higher scores were also common in central and western Moscow neighborhoods, which are generally considered more desirable.

The score also followed a sensible pattern based on the construction year. Some older apartments scored relatively well because they had been carefully renovated, while many apartments built during less prosperous Soviet periods scored lower. Newer apartments generally had high scores.

The score was not mainly measuring photograph quality

The researchers measured technical features of the pictures, including:

  • Sharpness.
  • Brightness.
  • Resolution.
  • Blurriness.
  • Shape and size of the image.

These technical features explained only a very small part of the differences in CLIP scores. This suggests that CLIP was mostly responding to what the apartment looked like, rather than simply rewarding bright or high-resolution photographs.

CLIP agreed with another AI system

The CLIP scores were positively related to the scores produced by LLaMA4, a multimodal LLM that can examine pictures and text.

This is important because the two systems use different approaches but reached similar conclusions about apartment quality.

The researchers argue that CLIP has some practical advantages:

  • It gives exactly the same result when run again.
  • It is inexpensive to use.
  • It can run on a local computer.
  • Images do not need to be sent to an outside company.
  • It uses relatively little computing power.

Higher-scoring apartments had higher prices

The CLIP Q-score was a strong predictor of both rental and sale prices.

After taking account of many other features, the researchers estimated that an apartment with a score 0.1 points higher would cost approximately:

  • 5.8% more to rent.
  • 3.5% more to buy.

Adding the CLIP score improved the accuracy of the price predictions. It was one of the more important features in the models. For rental properties, it was among the third most useful predictors. For sales, it ranked around fourth or fifth, depending on the model.

The relationship was not perfectly straight

The more advanced gradient-boosting model performed slightly better than the simpler regression model. It suggested that the connection between visual quality and price is not always exactly the same at every quality level.

For example, improving a very poor apartment may have a different effect from improving an already luxurious one.

Better-looking properties and time on the market

The paper reports that better-looking apartments initially appeared to stay online longer. This may seem surprising: shouldn’t attractive apartments disappear faster?

The researchers suggest an explanation. Better-looking properties usually have higher prices, and expensive apartments may take longer to rent or sell because fewer people can afford them.

To examine this, they used a proportional hazard model, a statistical method for studying the chance that something happens over time—for example, the chance that an advertisement is removed on a particular day. This model included both the CLIP score and the asking price.

The supplied text ends before giving the final detailed results of this analysis, but it shows that the researchers were trying to separate the effects of apartment quality from the effects of high prices.

5. Why are these findings important?

A normal real estate database may record facts such as size, location, and number of rooms. However, it may not record how modern, attractive, or well-maintained a home looks.

The CLIP Q-score can add this missing information by studying photographs. It works somewhat like an automatic visual inspector that can examine thousands of advertisements quickly.

The method may also be useful for other products. Researchers could change the three descriptions and use the same basic process to study:

  • Cars.
  • Hotel rooms.
  • Clothing.
  • Furniture.
  • Restaurant meals.
  • Houses in other countries.

For example, a car study might compare the descriptions “a high-quality car,” “an ordinary car,” and “a poorly maintained car.”

6. Simple conclusion

The paper shows that photographs contain valuable information about a product’s quality. A computer model called CLIP can turn that visual information into a simple number: the CLIP Q-score.

In the Moscow housing market, this score matched common ideas about which apartments were better, helped predict prices, and provided information that was not captured by ordinary property details.

The method could make online markets and economic research more informative. However, it should not be treated as a perfect or completely objective judge. The model learned from existing internet images and may reflect biases in those images or in the words used to describe quality. Also, a photograph can hide problems or make a property look better than it really is.

Overall, the research suggests that computers can use images to measure qualities that are difficult to describe with numbers—and that these visual qualities matter in real-life decisions about buying, renting, and selling homes.

View Paper Prompt  View All Prompts 

Knowledge Gaps

Knowledge gaps, limitations, and open questions

The paper leaves the following issues unresolved:

  • External validity across markets: It is unclear whether the CLIP Q-score performs similarly in other Russian cities, countries, housing markets, property types, or cultural contexts beyond Moscow apartments.
  • Temporal generalizability: The data cover only listings created from July 1 to August 14, 2025, with follow-up through December 5, so the stability of the score–price relationship across seasons, market cycles, and longer periods is unknown.
  • Reliance on asking prices: The analysis uses the last quoted sale or rental asking price rather than transaction prices, leaving unresolved whether the Q-score predicts actual negotiated prices or merely sellers’ pricing decisions.
  • Selection into online listings: The sample excludes properties not advertised on Cian.ru and removes listings without images. The extent to which image availability and platform participation are systematically related to quality, seller type, location, or price is not established.
  • Selection of images by advertisers: Advertisers may selectively photograph only the best rooms, omit defects, use professional staging, or upload misleading images. The score may therefore measure marketing presentation rather than the underlying physical quality of the property.
  • Dependence on prompt wording: The paper does not systematically test how alternative positive, neutral, and negative text descriptions affect scores. The robustness of rankings to synonyms, language, grammatical structure, and culturally specific notions of “luxury” remains unknown.
  • Dependence on the chosen CLIP model: Only one OpenAI ViT-B/32 implementation is used. Performance, reliability, and cross-market comparability across CLIP versions, model sizes, training datasets, and open-source implementations are not evaluated.
  • Lack of a human-annotated benchmark: Agreement with LLaMA4 is not equivalent to validity, because both systems may share biases or may respond to the same superficial visual cues. Direct comparison with expert appraisers, buyers, or large-scale human ratings is still needed.
  • Unclear construct validity: “Luxuriousness,” physical quality, renovation quality, visual attractiveness, and perceived status are treated as closely related, but the paper does not distinguish or measure these constructs separately.
  • Image-level versus property-level aggregation: The use of a simple average across images may be sensitive to the number of photographs, repeated views, room composition, and the inclusion of unusually good or poor images. Alternative aggregation methods are not compared.
  • Room and content composition effects: The score may be driven by whether an advertisement contains kitchens, bathrooms, living rooms, bedrooms, views, or building exteriors rather than by overall property quality. The paper does not decompose scores by room type or image content.
  • Sensitivity to image preprocessing: The effects of cropping, aspect-ratio distortion, resolution, compression, image ordering, duplicate images, watermarks, and image enhancements are not fully investigated.
  • Insufficient treatment of photo quality confounding: The residual adjustment explains only 2.5–3.0% of image-level score variation and uses linear controls. Nonlinear interactions between lighting, composition, staging, and CLIP scores may remain.
  • Potential omitted-variable bias in hedonic models: Unobserved renovation expenditures, furniture quality, views, building amenities, noise, exact micro-location, seller characteristics, and listing-agent effort may jointly affect both images and prices.
  • Causal interpretation of the price coefficient: The positive CLIP coefficient is predictive, but the analysis does not establish that improving a property’s visual quality or photographs would causally increase its market price.
  • Unresolved mechanism behind liquidity effects: The paper does not identify whether higher scores reduce search costs, increase buyer interest, signal unobserved quality, attract different customer segments, or simply correlate with pricing and seller behavior.
  • Endogeneity of initial asking prices in survival models: Initial price is likely chosen jointly with the listing’s images and unobserved property quality. Conditioning on it may not adequately resolve confounding in the estimated time-on-market relationship.
  • Questionable proportional-hazards assumption: The survival analysis does not report tests of proportional hazards, alternative duration specifications, competing risks, recurrent listings, or sensitivity to different definitions of listing closure.
  • Interpretation of listing removal: An advertisement being taken down may represent a sale, rental, withdrawal, expiration, relisting, or change in strategy. The extent of misclassification in the time-on-market outcome is not quantified.
  • Censoring and limited follow-up: Sales listings have substantial right-censoring, and the observation window is relatively short. Results may change with longer follow-up or alternative assumptions about censored listings.
  • Possible dependence among observations: Multiple listings, reposts, agencies, and properties within the same building may create clustered or duplicated observations. The random train–test split may therefore overstate out-of-sample performance if related listings occur in both sets.
  • Limited validation of out-of-sample performance: The random split does not test performance on future time periods, unseen neighborhoods, new buildings, different platforms, or entirely different cities. Temporal and geographic holdout validation is needed.
  • Uncertainty around predictive importance: The reported Shapley rankings may depend on the feature set, model specification, correlation structure, and evaluation sample. Confidence intervals or stability analyses for these rankings are not provided.
  • Comparison with stronger visual baselines: The paper does not compare the Q-score with supervised image models, standard computer-vision embeddings, image-quality indices, multimodal valuation models, or human ratings under a common evaluation design.
  • Calibration of the score: The polarity measure is bounded between −1 and 1 but is not calibrated to an interpretable quality scale or validated for consistency across different listing types and image sets.
  • Potential demographic and socioeconomic bias: CLIP was trained on internet data and may encode cultural, linguistic, racial, class, or aesthetic biases. The paper does not assess whether scores systematically disadvantage particular neighborhoods, architectural styles, household types, or renovation traditions.
  • Fairness implications for automated valuation: If the score is incorporated into pricing, ranking, lending, or housing-search systems, its effects on access, discrimination, and neighborhood sorting remain unexplored.
  • Robustness to manipulation: The paper does not test whether sellers can increase scores through staging, filters, lighting changes, wide-angle lenses, generated images, or carefully selected photographs without improving the property itself.
  • Privacy and governance questions: Although local inference improves data confidentiality, the paper does not discuss privacy risks associated with interior images, retention of scraped data, platform terms of service, or responsible deployment of the metric.
  • Incomplete assessment of LLM validation: LLaMA4 scores are available for only about 90% of ads, with up to ten randomly selected images for longer listings and one possible resubmission. The effects of missingness, truncation, random image selection, and provider-specific inference behavior are not examined.
  • Reproducibility beyond the scoring algorithm: Exact reproducibility may depend on the model checkpoint, tokenizer, preprocessing library, software versions, hardware, and text prompts. The paper does not provide a complete reproducibility protocol or sensitivity analysis for these components.
  • Transferability beyond real estate: The claim that the method generalizes to other products is not empirically tested. It remains unknown whether three ordinal text prompts are sufficient for products with different quality dimensions, brand effects, or functional attributes.
View Paper Prompt  View All Prompts 

Practical Applications

Immediate Applications

  • Real-estate automated valuation models (AVMs): Add the adjusted CLIP Q-score to existing hedonic, gradient-boosting, or comparable-sales systems to estimate rental and sale asking prices more accurately. The paper reports out-of-sample improvements in predictive performance and estimated price premiums associated with higher visual-quality scores. Sector: Real estate, proptech, finance. Dependencies: Local retraining and validation are needed for each city, property type, platform, and market period; the score should supplement—not replace—physical inspections and transaction data.
  • Listing-price recommendations for agents and property owners: Integrate the score into a broker dashboard that combines area, location, building characteristics, and image-derived quality to recommend an initial asking price. Potential product: An API or plug-in for real-estate listing software that calculates property-level visual quality from uploaded photographs. Dependencies: The reported relationship is based on asking prices in Moscow during a specific 2025 observation window, so it should not be interpreted as a causal or universally applicable price effect.
  • Automated listing-quality diagnostics: Use CLIP Q-scores alongside sharpness, brightness, aspect ratio, and BRISQUE metrics to flag listings whose photographs convey low perceived quality. Platforms could recommend additional photographs, better staging, or professional photography. Sector: Online marketplaces, marketing, software. Dependencies: The Q-score captures both property appearance and presentation context; technical image-quality controls remain necessary because image characteristics have measurable correlations with the score.
  • Search ranking and recommendation: Real-estate platforms can use the score as an additional ranking or filtering variable—for example, “visually upscale,” “renovated-looking,” or “best-presented”—to improve relevance for users who value appearance. Dependencies: Ranking systems should avoid systematically disadvantaging lower-income neighborhoods or properties with poorer photography. Users should be shown that the metric concerns visual presentation, not verified structural quality.
  • Listing comparison and portfolio screening: Buyers, renters, brokers, and property managers can use standardized scores to compare large numbers of listings before manual review. Investors could screen properties likely to command higher rents or sale prices. Dependencies: Scores are not substitutes for legal checks, building-condition surveys, energy audits, or neighborhood due diligence.
  • Marketing and staging evaluation: Agents can compute the score before and after renovation, decluttering, staging, or re-photography to assess whether visual changes improve the apparent market positioning of a listing. Potential workflow: Upload images → calculate per-image and property-level scores → compare alternative image sets → select the most informative listing package. Dependencies: A higher score may reflect luxury styling, lighting, or photographic composition rather than durable improvements to the property.
  • Real-estate market research and academic measurement: Researchers can reproduce the open method locally to construct image-based measures of housing quality, study hedonic pricing, analyze imperfect information, and augment traditional administrative or listing datasets. Sector: Economics, urban studies, marketing, computational social science. Dependencies: Researchers must document prompt wording, CLIP model version, preprocessing, image selection, aggregation, and adjustment procedures to ensure comparability.
  • Urban monitoring and housing-condition mapping: Municipalities and analysts can aggregate scores geographically to identify differences in visible housing condition across districts and monitor changes following redevelopment or renovation programs. Policy use: Prioritizing areas for inspection, façade improvements, housing renewal, or targeted surveys. Dependencies: Images are selectively uploaded, may be outdated or staged, and may underrepresent properties without photographs. Scores should not be used alone to allocate public services or determine eligibility.
  • Property-market liquidity analysis: Brokers and platforms can combine Q-scores with asking prices to estimate likely time on market and identify listings that are visually attractive but mispriced. The paper finds that visual quality is associated with market outcomes, particularly in sales after accounting for price. Dependencies: Time-on-market results are market-specific and affected by censoring, price strategy, seasonality, platform behavior, and whether an advertisement is withdrawn for reasons unrelated to a completed transaction.
  • Privacy-preserving visual analytics: Organizations handling sensitive property images can run CLIP inference locally rather than uploading photographs to an external multimodal-LLM provider. This offers a practical workflow for confidential brokerage, institutional housing, or research data. Dependencies: Local deployment still requires appropriate access controls, retention policies, copyright compliance, and assessment of model bias.
  • Cross-domain product-quality scoring: E-commerce catalogs, furniture marketplaces, vehicle listings, hotel platforms, restaurants, and rental services can adapt the method by replacing the apartment prompts with positive, neutral, and negative descriptions appropriate to the product. Example: “premium sofa,” “ordinary sofa,” and “worn-out sofa.” Dependencies: Prompt wording must represent the intended construct, and domain-specific validation is required because CLIP may associate quality with brand, style, demographic, or contextual cues.

Long-Term Applications

  • Multimodal valuation systems combining images, text, and video: Future AVMs could jointly model photographs, listing descriptions, floor plans, virtual tours, satellite imagery, street views, and transaction histories. This could distinguish visual polish from structural condition and improve robustness across property types. Dependencies: Requires larger labeled datasets, calibrated fusion methods, temporal tracking, and safeguards against duplicated or manipulated content.
  • Causal measurement of renovation returns: Researchers could use repeated images before and after refurbishment, together with sale or rental outcomes, to estimate the market value of specific visual improvements such as kitchens, bathrooms, lighting, or flooring. Dependencies: Strong identification is needed to separate renovation effects from broader market changes, seller selection, location, and changes in pricing strategy.
  • Fairer and more transparent property assessment: A validated, locally calibrated visual-quality index could support appraisal audits and reveal whether comparable properties are priced differently because of observable presentation quality. Explainability tools such as feature attribution and image-region analysis could show which visual elements contribute to predictions. Dependencies: CLIP can encode social and geographic biases. Deployment would require subgroup testing, human review, appeal mechanisms, and restrictions on using protected or proxy characteristics.
  • Public-housing and redevelopment policy evaluation: Governments could track image-derived changes in visible housing conditions before and after subsidies, demolition, rehabilitation, or neighborhood investment. This could complement inspections and resident surveys. Dependencies: Public policy decisions require representative data and cannot rely on marketplace images alone; image availability and staging may change over time independently of actual housing conditions.
  • Dynamic pricing and adaptive listing strategies: Platforms could develop systems that jointly optimize asking price, photo ordering, recommended staging, and publication timing. The objective would be to balance expected price against probability of sale or rental. Dependencies: This requires causal experimentation, real-time market data, and explicit modeling of strategic responses by agents and sellers. Optimization could also amplify inequality if not carefully governed.
  • Personalized visual search and consumer decision tools: Consumer applications could let users specify preferences such as “luxury,” “modern,” “well maintained,” or “needs renovation,” then rank listings using calibrated image-text similarity. Users might compare the visual quality of properties against price and location trade-offs. Dependencies: “Luxury” and “quality” are subjective and culturally variable; interfaces should present uncertainty and avoid implying that visual appearance guarantees safety, habitability, or construction quality.
  • Domain-specific, fine-grained quality scales: Instead of three prompts, future systems could use ordered descriptors—such as “unsafe,” “poor,” “basic,” “renovated,” “premium,” and “luxury”—or learn calibrated scores from expert and consumer judgments. Dependencies: Requires representative annotations, tests of rating consistency, multilingual prompt design, and validation against objective outcomes rather than agreement with another AI model alone.
  • Quality and defect detection for insurance, lending, and asset management: With further validation, image-based systems could help insurers, lenders, and institutional landlords triage inspection needs, monitor deterioration, or identify discrepancies between declared and visible property conditions. Dependencies: High-stakes use demands high sensitivity and specificity, secure data governance, resistance to adversarial or selectively framed images, and mandatory human inspection before underwriting or claims decisions.
  • Longitudinal urban and socioeconomic indicators: Aggregated visual scores could become inputs to urban-economics models measuring neighborhood investment, housing-stock aging, redevelopment, and spatial inequality over time. Similar methods could extend to storefronts, infrastructure, streetscapes, and public amenities. Dependencies: Longitudinal comparability is difficult because cameras, platforms, image conventions, and listing populations change. Calibration against censuses, inspections, and independent socioeconomic indicators is essential.
  • Standardized open benchmark for image-based product quality: The method could support a cross-sector research benchmark comparing CLIP variants, multimodal LLMs, image-quality controls, human ratings, and market outcomes. Dependencies: Progress requires shared datasets, licensing permissions, reproducible code, clear definitions of quality, and evaluation protocols that test geographic, cultural, and temporal generalization.
View Paper Prompt  View All Prompts 

Glossary

  • Adjusted residual: A regression residual modified to remove systematic effects of control variables. “the adjusted score is the average residual from the regression model”
  • Aspect ratio: The proportional relationship between an image’s width and height. “Aspect ratio is width/heightwidth/height.”
  • BRISQUE: A no-reference image-quality metric that assesses distortion using natural-image statistics. “The BRISQUE model compares the image's pixel distribution against that of a training dataset of pristine images.”
  • Censored observation: An observation whose event time is only partially known because the event has not been observed by the end of the study. “we face much higher right-censoring of durations”
  • CLIP: A multimodal model trained to align representations of images and text. “CLIP is a general purpose technique to train multimodal neural networks”
  • CLIP Q-score: A scalar image-based quality measure derived from the difference between CLIP’s positive and negative text probabilities. “Throughout the paper, we refer to the scalar polarity measure as the CLIP Q-score.”
  • Choropleth: A map that represents values across geographic areas using different colors or shading. “Finally, in figure~\ref{fig_clip_choropleth} we show how CLIP Q-scores vary across Moscow districts.”
  • Contextualized embedding: A vector representation whose meaning depends on the surrounding input sequence. “the models are capable of generating contextualized embeddings of their input sequences.”
  • Contrastive cross-entropy loss: A training objective that increases similarity for matching pairs and decreases it for mismatched pairs. “The symmetric contrastive cross-entropy loss is calculated as follows:”
  • Contrastive learning: A machine-learning approach that learns representations by distinguishing related pairs from unrelated pairs. “they proposed a loss function that would reward the models when they produced similar embeddings for image-text pairs that matched”
  • Cross-entropy: A loss function measuring the discrepancy between predicted probabilities and target labels. “The CLIP objective is operationalized by minimizing the average column-wise and row-wise cross-entropies”
  • Embedding: A numerical vector representation of an input such as text or an image. “The model's output from step 1 will consist of an embedding vector for the image”
  • Ensemble technique: A method that combines multiple predictive models to produce a stronger model. “Gradient boosting is an ensemble technique”
  • F-statistic: A statistic used to test the joint significance of multiple regression coefficients. “F-Stat & 321.2∗∗∗^{***}”
  • Gradient boosting: An ensemble-learning method that sequentially adds weak models to correct previous prediction errors. “we trained a histogram-based gradient boosting algorithm.”
  • Hazard function: The instantaneous event rate at a given time, conditional on the event not having occurred earlier. “where tit_i is the observed duration for property ii and h(⋅)h(\cdot) is the hazard function.”
  • Hedonic model: An econometric model that explains a product’s price through the implicit values of its attributes. “There is a long tradition of empirical investigations linking real estate values to observable characteristics.”
  • Histogram-based gradient boosting: A gradient-boosting implementation that discretizes continuous features into histogram bins for computational efficiency. “we trained a histogram-based gradient boosting algorithm.”
  • Hyper-parameter: A model configuration value selected before or during training rather than learned directly from the data. “Key hyper-parameters like the number of trees, the learning rate, and the tree depth”
  • Inference provider: An external service that runs a trained model to generate predictions. “the LLM approach requires submitting data to a service provider”
  • Kaplan–Meier estimator: A nonparametric estimator of a survival function that accounts for right-censored observations. “Kaplan-Meier right-censoring-adjusted survival curve estimates”
  • Log-linear model: A regression model in which the logarithm of the dependent variable is modeled as a linear function of predictors. “a common approach is to specify a log-linear model”
  • Logits: Unnormalized scores that are transformed into probabilities, commonly through softmax. “The CLIP objective is operationalized by minimizing the average column-wise and row-wise cross-entropies of the logits”
  • Laplacian filter: An image-processing operator that emphasizes regions of rapid intensity change, such as edges. “This method passes a Laplacian filter over the image to detect edges.”
  • Marginal contribution: The additional value attributable to adding one feature or participant to an existing set. “a weighted average of the marginal contribution of a feature”
  • Multimodal LLM: A LLM capable of processing text together with other modalities, such as images. “we compare the CLIP Q-scores with an alternative measure obtained from querying a multimodal LLM.”
  • Multimodal neural network: A neural network that accepts and jointly represents inputs from multiple data modalities. “deep models which can accept both images and text as inputs.”
  • Out-of-sample performance: Predictive accuracy evaluated on data not used to estimate or train the model. “Both the in-sample (training set) and out-of-sample (test set) performance”
  • Polarity score: A scalar sentiment-like measure formed by subtracting the probability of a negative category from that of a positive category. “take the polarity (P(+)−P(−)P(+) - P(-)) score as the final output.”
  • Pre-processor: Software that transforms raw inputs into the format and dimensions required by a model. “OpenAI has released a pre-processor that converts image-text batches into the correct format.”
  • Proportional hazard model: A survival-analysis model in which covariates multiplicatively scale a baseline hazard. “we estimated a proportional hazard model”
  • Right-censoring: The condition in which an event’s exact occurrence time is unknown because observation ends before the event occurs. “Kaplan-Meier right-censoring-adjusted survival curve estimates”
  • Shapley value: A cooperative-game-theoretic attribution score measuring a feature’s average marginal contribution across subsets of features. “the widely accepted approach to determining the relative importance of different predictive features is the Shapley value decomposition”
  • Similarity matrix: A matrix containing pairwise similarity scores between two sets of representations. “where S\boldsymbol{S} is the n×nn\times n similarity matrix”
  • State-of-the-art: Representing the best currently reported performance or methods in a field. “obtained remarkable state-of-the-art zero-shot performance”
  • Survivorship bias: Bias caused by analyzing only entities that remain observable or survive a selection process. “in order to avoid survivorship bias”
  • Temperature parameter: A scaling parameter that controls the concentration or randomness of a model’s output probabilities. “setting the temperature parameter to zero.”
  • Tokenization: The process of dividing input data into discrete units that a model can process. “It uses a ViT model that tokenizes 224×224224\times 224 pixel images into (49) 32×3232\times 32 patches.”
  • Transformer encoder: A neural-network component using attention mechanisms to convert an input sequence into contextual representations. “most current versions combine a regular transformer encoder”
  • Vision Transformer (ViT): A transformer architecture that processes images as sequences of image patches. “It uses a ViT model that tokenizes 224×224224\times 224 pixel images”
  • Zero-shot performance: A model’s ability to perform a task without task-specific training examples. “obtained remarkable state-of-the-art zero-shot performance on a number of benchmarked tasks.”
View Paper Prompt  View All Prompts 

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Generate Now