P3R:Polymodal Palpebral Progressive Refinement via Symmetry Aware Latent Diffusion for Precision Guided Prediction of Postoperative Blepharoptosis Morphology

Blepharoptosis is a descent of the upper eyelid that can impair vision. In surgical planning, clinicians need accurate, anatomically aware forecasts of postoperative appearance. Prevailing approaches such as geometric modeling, linear regression on margin reflex distance 1, and generative adversarial networks either oversimplify dermal and muscular dynamics or lack reliable anatomical control. We introduce Polymodal Palpebral Progressive Refinement (P3R), a framework that fuses three complementary priors, (i) concise, clinician verifiable text descriptors embedded via Contrastive Language–Image Pretraining, (ii) ocular masks derived from the Segment Anything Model that confine edits to operative regions, and (iii) a symmetry corrected contour prior that transfers healthy side geometry to the affected eyelid. A multi head cross attention mechanism aligns textual and visual cues during denoising to preserve eyelid topology and iris exposure. On 44 pairs from 30 subjects, Polymodal Palpebral Progressive Refinement improves anatomical indices over strong baselines while maintaining high image fidelity, thereby supporting precision counseling and surgical decision making. The results highlight the need for symmetry aware, multimodal conditioning to meet submillimetric clinical requirements.