PartLock: Vision–Language Model-Guided Part-Level Mesh Dragging

3D mesh dragging enables intuitive local shape editing by specifying a small number of handle points and their target positions, and has important applications in 3D content creation, animation, and industrial design. Existing diffusion-prior-based methods can improve the visual plausibility of large deformations, but they typically rely on global optimization without explicitly modeling semantic parts or editing regions, causing local manipulations to unintentionally affect unrelated regions. To address this issue, we propose PartLock, a vision–language model (VLM)-guided method for part-level mesh dragging. PartLock constructs a part-based proxy using PartField and determines the editable region through a three-stage VLM reasoning scheme, together with geometric validation to improve the completeness and reliability of the selected region. The resulting editable-region selection is then mapped back to the original mesh, where optimization is restricted to the selected editable region, while hard positional constraints are imposed on non-editable vertices. A diffusion prior is further employed for shape refinement. The experimental results show that PartLock can suppress unintended deformation propagation and produces natural deformation results while preserving the stability of non-editable structures. Moreover, it supports independent, asymmetric, and personalized part-level editing according to user-specified drag constraints.