MorphTwin: Discovering and Morphing Material Responses from Videos for Interactive Physical Twins

Reconstructing interactive physical twins from videos holds promise for virtual reality and embodied intelligence, but existing methods face a trade-off among simulation efficiency, material modeling accuracy, and explicit material control. Spring-mass models support efficient explicit simulation through simple discrete structures, providing an attractive foundation for real-time, interactive physical twins. However, their native spring parameters depend on graph structure, hindering cross-object reuse, while bulk-shear coupling in conventional central-force springs limits material expressiveness. Our key insight is to represent material behavior through standardized responses to volume and shape changes, separating a reusable material description from differences in spring geometry and connectivity across objects. Based on this insight, we propose **MorphTwin**, which augments explicit spring graphs with internally relaxable response carriers and performs offline energy calibration to compile shared material-response changes into graph-specific stiffness adjustments. This design supports material identification, interpretable editing, and cross-graph reuse while retaining efficient explicit dynamics. On synthetic scenes with known material parameters, MorphTwin reduces the mean absolute Poisson-ratio error and mean relative Young's-modulus error by 75.8% and 22.1%, respectively, compared with PhysTwin under the same inverse protocol. Real-world video experiments demonstrate improved reconstruction and future prediction and support reconstructing interactive physical twins from egocentric observations. With offline material calibration, online rollout advances the calibrated explicit spring graph without material identification or calibration optimization, preserving the real-time interaction capabilities of explicit spring-mass methods.