How it works The analysis Simulations Dashboard Lead generation via Google Ads Gallery See it for my clinic
Inside the engine

Why a client believes the analysis — and why the simulation still looks like them.

Anyone can send a photograph to an image model and call what comes back an analysis, or a simulation. It is usually confident, specific and wrong — which is worse than useless when your clinic's name is on it. This page is what the platform does instead, in enough detail to be checked.

Part one

The analysis

Four things stand between a photograph and an indicator a clinician would be comfortable standing behind.

  1. Every indicator is read several times, and the middle answer wins

    A single pass at an image model is not a measurement; it is one opinion with a confident voice. Each of the eleven indicators is read across several independent passes and the score is the middle of them, so one unusual answer cannot move the result. The spread is kept too — an indicator the passes disagreed about is treated differently from one they all agreed on. Below a minimum number of usable passes there is no analysis at all, rather than a confident-looking one built from a single sample.

  2. It is allowed to answer “I cannot see that”

    Each indicator is judged independently, and any one of them may come back unavailable with a short reason — glasses covering the under-eye area, hair across the forehead, lighting too flat to read texture. The rest of the face is still analysed normally.

    The instruction is explicit: a missing score is a correct and expected answer — an invented one is not. Most tools have no way to express this, so they return a number regardless, and nobody downstream can tell which numbers were real.
  3. A marker has to survive two separate tests before it is drawn

    A circle drawn on someone's cheek is a much stronger claim than a score. It says there, that one, look. So before any marker appears, two things must hold: a second independent read must have placed a marker in the same place, and the photograph's own pixels must support it — a spot marked as pigment has to actually be darker than the skin around it, a blemish has to be redder. That pixel test applies to the indicators drawn as individual points and lines, which is where a wrong mark does the most damage.

    This was built after measuring the alternative. On one face, six confident “dark spot” markers were checked against the photograph: five sat on skin no darker than its surroundings, and the sixth sat on an eyelash. All six now fail the test and are never drawn.
  4. A region it is unsure of is dropped, never nudged

    When the passes disagree about where something is, the overlay is omitted rather than placed somewhere plausible. A marker in the wrong place is worse than no marker at all, because it looks deliberate — and a client who spots one obvious mistake stops believing the other ten indicators too.

Part two

TrueRender™ — keeping the simulation true to the person

A simulation is only persuasive if the face in it is unmistakably theirs. Four mechanisms, each solving a specific way that image models fail.

  1. Identity lock

    Every simulation is built through a constrained prompt with a preservation block that cannot be edited away. Facial structure, jawline, chin, cheeks, nose, eyes, lips and proportions are held; so are age, ethnicity, hair, expression, lighting, background, camera angle and framing.

    It is still their face. That is enforced in code, not hoped for in a prompt.
  2. Sub-region rendering with feathered compositing

    Where a treatment affects one area, the engine does not regenerate the photograph. It locates the region, crops it, gives the model far more pixels on that area than a whole-face edit ever could, renders only that, re-aligns it, and composites it back through a feathered mask. Everything outside the region is the original photograph, pixel for pixel — it cannot drift, because it was never regenerated.

  3. Optical re-alignment

    Image models drift a few percent in zoom and offset, which is enough to make a comparison slider visibly jump. The engine estimates the shift by correlating greyscale thumbnails, applies the correction at full resolution, and refuses corrections above a threshold rather than faking them. Before and after line up, so the comparison is a comparison.

  4. A floor and a ceiling, running at once

    The floor: the result must always be an improvement — never duller, redder, older or worse in any way. The ceiling: show only what a skilled clinician could realistically achieve for this person in one reasonable course of treatment.

    The operative line is the last one: when uncertain, under-improve. A flawless result is not a better advertisement — it is one your client will not believe, or will believe and then hold you to.
Part three

What it will not do

These are not gaps waiting to be filled in a later version. They are the boundary that makes the platform safe for a clinic to put its name on, and they are enforced in the product rather than left to the copy.

Part four

What happens to the photograph

Consent first

Asked explicitly, in plain language, before a photograph is sent anywhere for analysis.

Deleted on a schedule

Face photographs are removed automatically once the retention period has passed. The job that does it is checked on every release, because a retention promise nobody verifies is not a promise.

Private links

A saved result opens only through a signed link, and those pages are never indexed by search engines.

Full detail in the privacy policy.

Run it on a photograph you bring.

Fifteen minutes, including the parts where it declines to answer.

Try it on a photograph