Restb.ai

Roughly a million property photos are uploaded every day in the US alone, by Restb.ai's own count. Every one of them contains information that the entire transaction depends on and that no system can read.

A human looks at a kitchen photo — the same image a tool like PhotoRoom would treat purely as something to clean up — and instantly knows the countertops are quartz, the appliances are recent, the finish level is above the neighbourhood average, and the whole thing was renovated in the last five years. A database looks at the same photo and sees a file called IMG_4471.jpg.

Restb.ai exists to close that gap. This review covers what the models actually output, why standardised condition scoring is the commercially important part, where it sits in valuation and MLS workflows, and what to interrogate before you feed a model's output into a number someone borrows against.

Property Photos As Structured Data

• Restb.ai is a computer vision platform that converts real estate photographs into structured, machine-readable property data — room type, features, architectural style, condition and quality.
• The condition and quality scores are the flagship. They map onto the standardised C1–C6 and Q1–Q6 scales appraisers already use, which is why valuation models can consume them directly.
• Founded in Barcelona in 2015, with a US presence, and an early €1.7 million award through the EU's Horizon 2020 programme.
• The customer list is the strongest signal here: Blackstone, Lemonade, idealista, Cotality, Black Knight, Stewart Valuation Intelligence and Utah Real Estate among others.
• It is infrastructure, not an app. You buy it to sit inside an AVM, an MLS, an appraisal platform or a portal — not to open in a browser.

Company, Models And Deployment

Why Standardised Condition Scoring Is The Whole Business

Of everything on the feature list, one capability carries the commercial weight: producing a condition and quality rating from a photograph on the same scale appraisers and lenders already use.

This matters because of how property valuation actually works. An automated valuation model can see square footage, bedroom count, location and sale history — all structured, all in public records. What it cannot see is whether the kitchen was gutted last year or last renovated in 1987. Two identical houses on the same street can differ enormously in value on exactly that axis.

That gap is why AVMs historically underperform on renovated and dilapidated properties specifically, and why the industry's answer was to send a human to look. A model that reads condition off the listing photos closes the gap without the site visit.

The company cites an AVM reducing its error rate by 9.2% using the condition models. That is a vendor figure and should be treated as one, but the direction is entirely plausible — condition is the largest unobserved variable in most residential valuation models.

The choice to output on the C1–C6 and Q1–Q6 scales rather than a proprietary score is the quietly clever part. Nobody has to learn a new vocabulary, and the output slots into forms, models and reviews that already exist.

The reason this is infrastructure and not a novelty: condition is the biggest thing a valuation model cannot see, and it is sitting in plain view in the listing photos.

The Second Job: Keeping Listing Photos Honest

A separate group of models does something less glamorous and, for an MLS, arguably more urgent — policing the images themselves.

Multiple listing services have rules about photographs. No agent branding in the image, no watermarks, no company logos, no photos of the wrong property, no duplicate images padding a listing. Historically these rules were enforced by complaint: someone noticed, someone filed, staff reviewed.

Watermark and added-logo detection, plus exact and near-duplicate identification, turn that into a check that runs on upload. That is a meaningful staffing change for an organisation processing thousands of new listings a week.

Restb.ai reads what is in a photograph; tools like REimagine Home work in the opposite direction, changing what the photograph shows. Duplicate detection has a second use worth naming. Near-duplicate matching across listings is how you notice the same photographs appearing on properties they do not belong to — useful for data quality, and useful for spotting listings assembled carelessly.

• Watermark and artificially added logo detection for MLS image rules
• Exact and near-duplicate identification across a listing and across the database
• Photo compliance monitoring against an MLS or site's own published guidelines
• Room and space classification so images can be ordered and labelled automatically
• Automated captions, which serve both search visibility and screen reader accessibility

Captions, SEO And Accessibility In One Move

Automated image captioning appears on the feature list, and it is doing two jobs that are rarely solved together.

The accessibility job is straightforward. A property listing with forty images and no alternative text is unusable with a screen reader, and image accessibility is a live compliance question for consumer-facing property sites.

The search job is that image alt text and surrounding descriptive copy are among the few remaining places on a listing page where a portal can add genuinely unique content. Every portal has the same address, the same square footage and frequently the same agent-written blurb. Descriptions generated from what is actually visible in each photo are unique per listing by construction.

The company cites a real estate portal seeing a 46% increase in Google traffic after adding SEO image captions. Again, a customer-reported figure rather than an independently audited one — but the mechanism is sound, and it is the same reason unique product copy outperforms manufacturer boilerplate in e-commerce.

A Blackstone subsidiary is cited as saving over $1 million annually through automated descriptions, which reads as a labour substitution figure rather than a revenue one.

Where This Sits In A Stack You Already Have

This is not a product you log into. It is an API that partner platforms embed, and the named partner list tells you more about the deployment model than any feature page.

Cotality — formerly CoreLogic — Black Knight, Stewart Valuation Intelligence and Bradford Technologies are all infrastructure vendors to the mortgage and valuation industry. If Restb.ai's models are reaching lenders, they are largely reaching them through those platforms rather than directly.

Lemonade points at the insurance path, where the same condition and feature extraction supports underwriting and claims. idealista points at the portal path in Europe. Utah Real Estate is an MLS. Blackstone is an institutional owner at scale.

The practical implication for a buyer: check whether the platform you already use has this integrated before scoping your own build. Buying it twice is a real risk in a market where the same vision layer is licensed widely.

For anyone building their own valuation or portal product, the questions are the ordinary ones for a vision API — throughput, latency, per-image cost structure, how confidence is expressed, and what happens on an image the model cannot classify.

1. Assemble a hundred properties where you already know the true condition and quality grades.
2. Deliberately over-sample the hard cases: partial renovations, dated but immaculate homes, and staged photography.
3. Run the images and compare the model's grades against your ground truth, tracking direction of error rather than just accuracy.
4. Check what the model returns when it is uncertain — a wrong grade delivered confidently is far worse than a flagged abstention.
5. Re-run the same set six months later to see whether the model has drifted under you.

The Cautions Worth Holding Onto

Three, and none of them are specific to this vendor — they apply to any vision model feeding a financial decision.

Photos are marketing artefacts, not evidence. Listing photography is deliberately flattering: wide lenses, staged rooms, careful angles, and increasingly virtual staging. A model reading condition from a marketed image is reading a curated version of the property, and the gap widens exactly where the stakes are highest.

Automated grades that influence lending carry fair-lending exposure. If a condition score moves a valuation and the valuation moves a loan, the model is inside a regulated decision. Testing for systematic differences in grading across neighbourhoods is not optional diligence.

The published outcome figures are customer-reported. A 9.2% error reduction and a 46% traffic lift are both entirely believable and neither is independently audited. Run your own pilot on your own data before either number appears in your business case.

The Vocabulary This Category Assumes

Real estate valuation has its own shorthand, and the product only makes sense once these are clear.

AVM

Automated Valuation Model — a statistical model that estimates a property's value from data rather than a human inspection. The main consumer of condition scores.

C1–C6

The standardised condition scale used in appraisal practice, running from newly built and never occupied down to substantial deferred maintenance.

Q1–Q6

The parallel quality-of-construction scale, describing finish and materials rather than wear — a home can be high quality and poor condition simultaneously.

MLS

Multiple Listing Service — the regional database agents list into. MLSs set the photo rules that compliance models enforce.

Comparable

A recently sold similar property used to justify a valuation. Visual similarity matching is an attempt to pick comparables on how the homes actually look, not just their attributes.