PartRM: Modeling Part-Level Dynamics with Large Cross-State Reconstruction Model

Figure 1. We present PartRM, given a single-view image and user-specific drags, PartRM can efficiently models appearance, geometry, and part-level motion in a feed-forward manner. Unlike the previous state-of-the-art, Puppet-Master [24], PartRM achieves higher PSNR and significantly faster inference times. As shown in (b) compared to (a), PartRM produces 3D-aware, part-level motions with enhanced multi-view consistency, delivering more realistic and coherent results across different viewpoints.