FusionDeformer: text-guided mesh deformation using diffusion models

Abstract:

Mesh deformation has a wide range of applications, including character creation, geometry modeling, deforming animation, and morphing. Recently, mesh deformation methods based on CLIP models demonstrated the ability to perform automatic text-guided mesh deformation. However, using 2D guidance to deform a 3D mesh attempts to solve an ill-posed problem and leads to distortion and unsmoothness, which cannot be eliminated by Clip-based methods because they focus on semantic-aware features and cannot identify these artifacts. To this end, we propose FusionDeformer, a novel automatic text-guided mesh deformation method that leverages diffusion models. The deformation is achieved by Score Distillation Sampling (SDS), which minimizes the KL-divergence between the distribution of rendered deformed mesh and the text-conditioned distribution. To alleviate the intrinsic ill-posed problem, we incorporate two approaches into our framework. The first approach involves combining multiple orthogonal views into a single image, providing robust deformation while avoiding the need for additional memory. The second approach incorporates a new regularization to address the unsmooth artifacts.

Our experimental results show that the proposed method can generate high-quality, smoothly deformed meshes that align precisely with the input text description while preserving the topological relationships. Additionally, our method offers a text2morphing approach to animation design, enabling common users to produce special effects animation.

bibtex:

@article{xu2024fusiondeformer,
  title={FusionDeformer: text-guided mesh deformation using diffusion models},
  author={Xu, Hao and Wu, Yiqian and Tang, Xiangjun and Zhang, Jing and Zhang, Yang and Zhang, Zhebin and Li, Chen and Jin, Xiaogang},
  journal={The Visual Computer},
  volume={40},
  number={7},
  pages={4701--4712},
  year={2024},
  publisher={Springer}
}