Title: Mastering Neural Deformation Representation for the Creation of 4D Dynamic Shapes

Abstract: The evolution of 3D shape representation has unlocked novel avenues for producing intricate 3D models. Nevertheless, research into the synthesis of 4D dynamic shapes—defined as 3D objects undergoing deformation across time—remains scarce. This study aims to fill that void by prioritizing both the fidelity and computational efficiency of 4D dynamic shape generation. Prior to this work, HyperDiffusion attempted to address 4D generation by directly synthesizing weight parameters for 4D occupancy fields; however, this approach was hindered by sluggish rendering speeds and poor temporal consistency, largely because the motion and shape representations were inextricably linked. To overcome these limitations, we introduce a novel neural deformation representation. By integrating this with conditional neural signed distance fields, we construct a 4D architecture that successfully decouples the motion latent space from the shape latent space. Our proposed method predicts rigid transformations and skinning weights for distinct segments, offering superior structural comprehension compared to the deformation modules found in current 4D representations. Furthermore, we have engineered a diffusion model training framework that leverages shape and motion features, derived from our 4D representation, as input data points. Experiments encompassing unconditional generation, conditional generation, and motion retargeting reveal that our approach outperforms existing methods in 4D dynamic shape generation and opens the door to a wide array of potential applications.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC