Our simulation engine is constrained by peer-reviewed anatomical principles, published injection frameworks, and clinical ethics standards, not aesthetic trend data.
Our approach
Med Plasty treats AI image generation as a constrained optimisation problem: the output must simultaneously satisfy anatomical plausibility, photometric realism, identity preservation, and clinical ethics.
Rather than fine-tuning on before/after clinical photographs (which would encode practitioner variance and operator bias), we encode the constraints directly into the prompt system using clinically authoritative terminology, injection plane names, rheology classifications, established morphometric references, that the model has learned from the medical literature it was trained on.
The goal is a simulation a board-certified aesthetic injector would be comfortable presenting to a patient during a regulated consultation, not a consumer-grade transformation.
Scientific foundations
The clinical authority signals encoded into every simulation prompt.
The global injection mapping standard used in Allergan/AbbVie certification worldwide. Our simulations name the points they use: Ck1 (lateral zygoma) and Ck2 (malar eminence) for cheeks, with Ck3 (anterior medial cheek) kept free of product; T1 (anterior temple) for temples; Jw1 (mandible angle) and Jw3 (mandible body) for the jawline; NL1 (upper fold), NL2 (central fold) and NL3 (lower fold) for the nasolabial folds; Tt1 (central infraorbital), Tt2 (lateral infraorbital) and Tt3 (medial infraorbital) for the tear trough; N3 (frontonasal angle), N4 (bony dorsum) and N5 (cartilaginous dorsum) for the nose, with N1 (anterior nasal spine) and N2 (columella) kept free of product.
The deep injection plane for structural cheek, chin, and jaw work. Prompts are named by the plane, not the surface landmark, so the model anchors on real anatomical geometry rather than visual approximation.
High G-prime HA (Voluma/Lyft class) signals structural, not soft, filler behavior. The simulation models lift versus spread based on rheological class, producing distinct visual outcomes from, say, a labial filler versus a malar bolus.
The Ogee curve is the universal aesthetic reference for mid-face contour taught in aesthetic fellowships globally. The Ricketts E-line is used for chin projection assessment. Both anchor the visual outcome in established morphometric standards.
MD Codes™ is a trademark of Dr. Maurício de Maio. Med Plasty is not affiliated with Dr. de Maio or Allergan/AbbVie.
Point names follow de Maio M, “MD Codes™: A Methodological Approach to Facial Aesthetic Treatment with Injectable Hyaluronic Acid Fillers”, Aesthetic Plastic Surgery, 2021 (PMID 32445044).
Prompt architecture
Each of the 15 supported treatments has dedicated prompts at Subtle, Moderate, and Dramatic intensities, with explicit clinical outcome descriptions, anatomical change specifications, visual outcome standards, and a comprehensive list of forbidden failure modes.
Lip augmentation additionally uses a continuous ML slider (0.5–5.0 ml), where the instruction text scales with dose, 0→1 ml is described as major change on thin lips; 4→5 ml as clinical ceiling for a single session.
Output verification
A rendered image is never trusted on sight. Before a doctor sees it, the system compares it with the patient's own photo. Anything that does not match is realigned, re-rendered, or replaced with the original pixels.
One by default. A second render runs automatically only when the first fails the framing check, and the better-framed of the two is kept.
The score a result gets against the original photo when the model keeps the framing, measured on real simulations.
Below this score the model has re-framed the photo, for example zoomed into the face. The result is not colour-corrected and the image is rendered again once.
Prompt version, model version, finish reason and number of model calls are stored for every simulation, so any result can be traced and compared later.
The checks, in order
The result is compared with the original across a search window of ±8% and shifted back into register, so the face sits exactly where it did in the photo.
A normalised similarity score shows whether the model kept the original framing. If it did not, the image is rendered once more and the closer match is kept.
Brightness and colour are pulled back toward the original within strict limits, so a change in lighting cannot pass as a treatment effect.
In the viewer the result is registered on the eyes and masked to the face. Hair always, and any eyes or nose the treatment does not target, come from the original photo. PRF changes only the skin under the eyes.
Each change is rendered at least twice, always from the patient's original photo and never from an earlier result. It is compared with the previous version on the same input, and measured where a feature can be measured, such as tooth height.
These checks catch framing, alignment and colour errors. They cannot judge clinical realism, and they do not guarantee an outcome. Every simulation is a visualisation for consultation, reviewed by the practitioner.
Academic inquiry
We welcome inquiries from academic institutions, aesthetic medicine researchers, and clinical safety organisations interested in our methodology, prompt architecture, or outcome validation.
Get in TouchBuilt on clinical science, evaluated against published standards, and constrained by the ethics of aesthetic medicine.