Certifying a pipeline, not a photo.

Pipeline certification under VQS, the Visual Quality Score published by Standard Vision Lab as an objective standard for real estate photography image quality, measures a production workflow rather than a single shoot. Standard Vision Lab chooses the input images, the applicant cannot preview, filter or cherry-pick them, and the images must cover a spread of room types and shooting conditions. It is the certification to look at when the claim being made is that a process is consistent, rather than that one job came out well.

What a set certification proves, and what it does not

A VQS run measures the files in front of it, and nothing else. A run takes 12 to 20 original JPEGs from a daytime shoot. Every image is fingerprinted with SHA-256, and duplicates within the set or against earlier runs are removed. Images score 0 to 100, with statuses applied per check and per image: PASS is 90 to 100, FLAG is 70 to 89, FAIL is 69 and below. The checks sit in families such as exposure, white balance, colour and geometry.

A set is not scored on its mean. It takes the score of its lowest image. Some faults are gates rather than deductions, which means they fail the set outright whatever the composite says, and focus is gated that way currently. One 19-image set had every image passing individually, a mean of 98 and a lowest of 93, and it still failed because two frames were soft. If that logic is unfamiliar, a set is only as good as its worst image covers it in full.

That result is precise, and it is narrow. It says those files met the standard at that version. It says nothing about the next job.

Pipeline certification measures the process instead

The standard treats pipeline certification as a distinct thing. It certifies a production workflow rather than one shoot.

That is the difference between showing a client a good shoot and telling a client what your process will do with work nobody has seen yet. A set certification cannot carry the second claim, because a set certification only ever describes the set that was submitted. Pipeline certification is built for it.

SVL chooses the inputs

The defining rule is short. SVL selects the images that go through the pipeline, and the applicant cannot preview, filter or cherry-pick them.

The reason is straightforward. A test you curate measures the work you chose. A test you cannot curate measures the process that handled it. A curated sample would say nothing about the work as a whole.

Blind selection also removes the awkward conversation. Nobody has to take your word for how the sample was assembled, because you did not assemble it.

Why the spread across room types and conditions matters

The inputs must cover a spread of room types and shooting conditions. That requirement does the same job as blind selection, approached from the other direction.

Set consistency is measured in its own right, separately from the composite. VQS reports sigma statistics per grouping, meaning global, interiors, exteriors and room class, together with per-image z-score outlier attribution that names the frame dragging the set down. Groupings need something in them before they can say anything, and a set drawn from one kind of room in one kind of light leaves those comparisons with very little to compare.

The checks themselves are named and citable, among them exposure_bright, exposure_dark, shadows_crushed, white_balance, white_balance_tint, white_balance_windows, colour_cast, contrast, saturation_under, grass_saturation, greens_saturation, wood_saturation, geometry_vertical, geometry_horizontal, noise, window_clipped, window_dark, window_artifacts, ghosting_artifacts, ai_texture_blacks, camera_reflection and cables_visible. The checks covers them. A workflow can be steady on one class of image and unstable on another, and a narrow input set is the easiest way for that difference to go unrecorded.

How the image was made is not an input to the score

VQS applies to imagery however it was produced, and a conforming runner must not use any declaration of production method as a scoring input. That cuts both ways, and both ways are useful. You are not marked down for your method. You also cannot claim credit for it. The output is measured, and only the output. AI-edited real estate photos takes that further.

It is worth being plain about the boundary. VQS measures technical execution. Composition, styling, aesthetic preference, storytelling, video, floor plans, listing text accuracy and legal compliance are out of scope. Portals publish their own technical upload requirements such as dimensions and file formats, agencies keep internal style guides, and editing suppliers work to commercial service agreements. None of those are what a VQS score is reporting on.

What determinism does and does not promise

The runner is deterministic: the same image gets the same score. That is a guarantee about the measurement, not a guarantee about your commercial claims, and two things sit outside it.

First, a score states the standard version it was produced under and remains valid against that version, so a version change can move numbers without anything in your work changing. Second, a pipeline test runs on images SVL selects, so a later test is not a rerun of an earlier one on the same files. Determinism means the instrument is stable. It does not mean two pipeline results are a like-for-like comparison.

What a material change does to the certification

If the pipeline materially changes after certification, for example swapping or retraining a model, the certification no longer covers output made after that change until the pipeline is re-tested.

Read that as a release-management rule rather than a penalty. It is what the certification is anchored to. If you cannot say when your pipeline last changed, that is worth fixing whether or not you ever certify.

It also protects the value of the mark. A certification that survived silent model swaps would be worth nothing to the people relying on it. The badge is a live widget rather than a static image, it links to a public verification register, and expired, suspended and revoked marks visibly say so.

What certification looks like when it lands

Certification is automatic. A certifiable composite earns its tier badge with the result: Platinum at 100, Certified at 90 to 99, Provisional at 70 to 89, no mark below 70, and no mark at all for a set that fails a gate. There is no application and no review panel. The badge is one line of HTML and it clicks through to a public register showing tier, composite, covered files and dates, and opening the full annotated report. It expires 12 months after the last certified run. How VQS certification works sets out the tiers in detail.

Pricing is US$9.99 a month, which includes one run each month and keeps the mark live while the subscription runs, with a recertification run required at least every 12 months, or US$129 for a single run with the badge valid for 12 months. Single-image certification is also defined, binding to one exact file by its SHA-256.

A sensible starting point

Run a set first. One submission returns a composite, per-check statuses and the consistency statistics for the files you sent, on work you selected yourself. If those hold up, a blind input set is a reasonable next step. If they do not, you have found the thing to fix before anyone else picks images on your behalf. There is a demo, and the standard explained if you want the shape of it first.

Certification is issued only by SVL, scored by its canonical runner. Scores produced by other conforming runners are advisory. The standard is versioned, so a score states its version and remains valid against that version. The full specification text is in preparation for publication and is available on request.