What the evidence says about AI before-and-afters

Who this is for

This is for the physician or medical professional who owns a cosmetic surgery or injectables practice and has been pitched a tool that draws a patient's likely result, and/or the operations lead who would put that tool in front of patients.

Surgeons have altered patient photos for decades. Canfield's Mirror software morphs them in consultation rooms today, and there was Photoshop and a pencil before that.

The mockup now looks like a photograph. The question is whether that change does anything for the decision a patient is making.

Do patients decide on the render?

The render came third, behind the practice's photos of real patients and behind the surgeon in the room.

A 2026 study in Computer Assisted Surgery followed 75 women through rhinoplasty, breast augmentation, mastopexy, augmentation-mastopexy, and reduction, all of whom received preoperative 3D simulations. A year after surgery, researchers asked what had driven the decision.¹

Photos of previous patients took 23 of the 75. Talking to the surgeon took 22. The simulation took 14. The remaining sixteen picked something the study doesn't list.¹

The top two have something in common. Photos show what a surgeon produces, and conversation shows whether that surgeon understands what a patient wants. Both carry information about the person who'll be holding the instruments.

The simulation carries none. It shows a possible result and says nothing about who would produce it.

The gap between the simulation and the consultation is 8 people (22 against 14), and the authors call their own study preliminary.¹

Patients arrive at this decision nervous, which is part of why the images matter. Asked to rank 11 kinds of content, 100 aesthetic patients put before-and-after photos first, in a 2017 American Society of Plastic Surgeons survey.²

A 2013 conjoint analysis of 150 patients, published in Plastic and Reconstructive Surgery and released by the American Society of Plastic Surgeons, broke the choice of surgeon into weighted attributes. A conjoint analysis asks people to trade attributes off against each other, which shows how much each one counts.³

Referrals from a GP, the primary care physician, carried weight. Referrals from TV, radio, and internet forums had a negative effect.³

We sell digital marketing, and the online layer in this table came to 9%. Surgeon experience, at 36%, was 4 times that. A surgeon's reputation outranks anything we can build for a practice.³

A website simulator works on online presentation, the second-smallest of the six attributes, and gets sold as a differentiator.

Which choices can a render help a patient make?

Renders are best at comparing options a patient and surgeon get to choose between.

Where an outcome is largely set by something the surgeon and patient select, simulation has material to work with: filler volume and placement, implant size and profile, a lift alongside augmentation versus augmentation alone, or a choice between two feasible reduction targets.

The evidence is specific about where this applies. In the same 2026 study, patients rated how much the simulation influenced them on a 1 to 10 scale. Breast augmentation scored 8.4. Breast reduction scored 3.8.¹

Same body part, and the gap tracks how much of the result can be picked off a shelf. Augmentation is largely set by an object with known volume and projection. Reduction is set by tissue quality, pedicle choice, and how the scars behave.

From those two data points we draw a working rule of our own, which the literature does not state. Simulation is most useful where the result is set by a choice, and least useful where it is set by how tissue heals.

The clearest published example gets this right. A 2025 pilot at MedStar Georgetown, published in Plastic and Reconstructive Surgery Global Open, used soft tissue simulation to help seven breast reduction patients choose between two feasible cup sizes. The authors cite earlier work putting the gap between perceived and measured cup size as high as 70 to 100%.⁴

Patients rated that choice 9.6 out of 10 for difficulty, and every respondent called the images extremely helpful. The authors framed it as a decision aid and said plainly it wasn't a prediction.⁴

The study had seven patients, and nobody has replicated it since.

How close does a render come to the healed result?

A prediction can only be checked once the patient it was made for has healed. That window runs weeks for a neuromodulator, a few months for filler or a laser series, several months for breast surgery, and 12 to 18 months for a nose.

So validation has to be judged procedure by procedure. A filler simulator that launched in 2024 has had time to compare its predictions against real outcomes. A rhinoplasty simulator that launched the same year has not.

The marketing also blurs training with validation. Training a model on historical cases is a different thing from checking what it predicts. EntityMed says its facial simulator draws on more than 50,000 clinical cases, which describes training data. It says nothing about whether any render was later held up against how that patient healed.⁵

Where accuracy has been measured, the results are approximate, and the numbers are not comparable with each other, since each was produced by a different method on a different procedure. In the 2026 study, patients rated the resemblance between render and result at 7.5 out of 10 for rhinoplasty and 7.9 for breast augmentation; that is patient self-assessment, with no blinded measurement.¹

A separate 2025 peer-reviewed study of 38 rhinoplasty patients simulated on Crisalix found accuracy correlated with satisfaction at 0.66. A 2025 prospective cohort of 78 breast augmentation patients imaged on Canfield's VECTRA found the predicted volume within 10% of the result in 73% of cases.^6,7

Open versus closed rhinoplasty is an access strategy, and the external result can be identical either way. The differences sit in the exposure, the dissection, the grafting, and the revision profile. A render of the outside of a nose shows none of it.

Septoplasty versus septorhinoplasty is sharper still, because septoplasty is done for airflow. No vendor or published workflow we found compares two surgically distinct techniques on the same patient.

Nobody has tested whether photorealism does anything. There is no published experiment comparing photorealistic renders against rough morphs on how firmly patients hold onto them. Looking more real is the entire commercial argument, and it is unmeasured.

The nearest published work is not encouraging. A 2026 systematic review in npj Digital Medicine found clinical fidelity problems in 47.2% of the studies of AI-generated medical images it examined.⁸ A 2025 peer-reviewed review of AI beauty filters found believable changes to nasal tip rotation, lip fullness, and jaw prominence produced while anatomy gets ignored.⁹

Who sells these tools, and what do they claim?

Sixteen products generate a patient-specific outcome image. They fall into three groups by where the image is delivered: the consultation room, the practice website, and apps patients install themselves. Practices are sold the first two.¹⁰

Three more names, RealFaceValue, New Look Now and AesthetIQ AI, have no current product documentation or turned out to be something other than a simulator. Canfield Mirror, oVio360 and TouchMD sit outside the table on purpose, since none of them generates a prospective image.

The first group sits in the room with a clinician. The output gets discussed in the room instead of handed over, which suits an approximation.

Allergan publishes an 86% consultation-to-treatment data point for Crisalix on its Natrelle materials, sourced to physician user data.¹¹ Johnson & Johnson publishes 87% for the Mentor-branded Arbrea app, sourced to a conference abstract.¹²

Two competitors landed a percentage point apart, neither data point rests on a study, and each tool shows its own sponsor's implants. Device makers have always funded the tools surgeons use, and that 86% came from the implant maker.

The second group never meets a clinician. It generates a preview for an anonymous website visitor and sells that to practices as lead generation, and it's priced accordingly. PreviewMD's advertised plans charge the practice a dollar per preview past plan, which means the vendor's revenue scales with the number of images generated, whatever their quality.¹⁰

Render time separates the two tiers cleanly. The MedStar decision aid takes a week or more.⁴ PreviewMD takes three to four minutes, and ClinicOS advertises under five seconds.¹⁰ The Mentor app captures and simulates in sixty seconds.¹² The extra time is clinician review.

Basic diligence on the website tier mostly came up empty. Across that group, sites often named no legal entity, no founders, no launch date, and no underlying image model. Where a site claimed photos are never stored, a privacy policy or data processing agreement to check that claim was often missing.¹⁰

Some of this group also assert in their own marketing that no business associate agreement is needed. Whether one is needed is a determination for the practice and its counsel, and it is not one a vendor gets to make on the practice's behalf in a FAQ.

Most of what a simulator gives a practice is a signal to patients. This is our judgment, with no study behind it: the pitches lead with differentiation and lead generation. If the value were clinical accuracy, a buyer would expect accuracy data in the marketing, and what the marketing offers instead is conversion data.

Signaling that a practice invests in technology is legitimate. Buying a VECTRA says something about the surgeon who owns it, the same way a renovated waiting room does. The difference is that the practice captures the benefit of looking current while the patient carries an image of their own face implying a result nobody validated.

What should a practice ask before it buys a simulator?

Which tier it is. In the consultation room, a simulator can be used the way the studies used it, showing two or three options side by side; none of them studied it as a prediction.^1,4

For a website tool, where the photograph goes, who processes it, and who decided no business associate agreement was needed; a practice already running one can read its patient-facing page the way patients and the FTC do, headline, imagery and disclaimer together.

For any vendor, who produced the conversion number and how conversion was defined, whose implants are in the catalog, and whether "validated" means training data or a comparison against healed results, procedure by procedure.

On a finite budget, photos of real patients outranked simulation as a decision factor and cost nothing beyond the discipline to shoot them consistently.¹

About this review

The review covered sixteen products identified and priced from public vendor materials, plus peer-reviewed literature on simulation accuracy and decision-making retrieved through August 2026, plus public regulatory guidance. Manufacturer-published and vendor-published data points are labeled as such throughout and were not treated as independent validation. In October 2026 every source was re-found and given a retrievable reference; one could not be, and its data points were removed (see 7.2).

The full text of the 2026 simulation study was not accessible, so the survey instrument is inferred: the reported shares resolve to whole patients out of 75, which indicates a single-select question, and the fourth category in the chart is derived by subtraction; the authors did not report it. Pricing is as advertised and was not independently confirmed. There is no public market-share data for this category. A 2025 survey of 122 patients on preoperative concerns, cited in the August version, could not be retrieved in October 2026 and has been removed. And nobody has compared photorealistic renders against rough morphs on expectation anchoring, which is the most important open question.

Fractional Aesthetics is a growth strategy practice for aesthetic medical businesses.