Toward Clinically Interpretable Postoperative Prediction: A Diffusion-Based Framework with Unsupervised Geometric Priors

Predicting postoperative outcomes is crucial for effective communication and planning in aesthetic medicine; however, current approaches often rely on manual measurements performed by clinicians. We present an annotation-free pipeline that integrates unsupervised geometric priors with diffusion-based inpainting to visualise postoperative results. Structural constraints are derived from edge-based features within a standardised periocular region of interest, avoiding reliance on segmentation or landmarks while ensuring anatomical plausibility. We evaluate the pipeline on 131 paired preoperative–postoperative eyelid surgery samples under a zero-shot, cross-domain setting without fine-tuning or retraining. Quantitative results demonstrate strong performance across three dimensions: identity preservation, boundary consistency and perceptual quality. Qualitative analyses further confirm that the model produces natural, continuous creases aligned with postoperative morphology, while maintaining subject-specific identity and seamless blending. These findings highlight the feasibility of combining geometric priors with generative diffusion models for the prediction of interpretable and reproducible outcomes. While some failures remain in underrepresented subgroups, the proposed zero-shot, annotation-free workflow demonstrates potential as a lightweight tool for surgical visualisation, with applicability beyond eyelid surgery to broader aesthetic and reconstructive domains.