Neural Clustering for Prefractured Mesh Generation in Real-time Object Destruction

TL;DR

This work tackles the realism-versus-performance trade-off in real-time object destruction by learning to cluster prefractured meshes based on structural weakness. By treating prefracture generation as unordered clustering on a center-of-mass point cloud and employing a Point Transformer backbone with a permutation-invariant loss, the method predicts meaningful groupings. Trained on a physics-based Breaking Bad dataset and evaluated on unseen Objaverse data, it produces ready-to-use prefractured meshes that better reflect weak points than baselines. The approach enables more realistic destruction in real-time engines such as Unreal Engine 5, potentially reducing manual artist intervention and improving visual fidelity.

Abstract

Prefracture method is a practical implementation for real-time object destruction that is hardly achievable within performance constraints, but can produce unrealistic results due to its heuristic nature. To mitigate it, we approach the clustering of prefractured mesh generation as an unordered segmentation on point cloud data, and propose leveraging the deep neural network trained on a physics-based dataset. Our novel paradigm successfully predicts the structural weakness of object that have been limited, exhibiting ready-to-use results with remarkable quality.

Figures (1)

• Figure 1: Proposed network architecture. (a) Point Transformer zhao2021point as a backbone network. (b) The softmax of output as the probability that each point belongs to each group, where $N=|X|$ and $K=|G|$. (c) Matrix multiplication of (b) derives point-to-point group identity, providing permutation-invariant loss to group.