上海科技大学知识管理系统

We present 3DMorphing, which aims to generate smooth and reasonable 3D interpolations between two 3D objects. Powerful 3D generation and reconstruction struggle to achieve this task due to the highly unstructured diffusion latent space and the non-semantic feature space, respectively. 3DMorphing utilizes the prior knowledge of multi-view diffusion models to achieve smooth and reasonable 3D interpolation. Our core idea is to use two LoRAs to fit two 3D objects to capture their respective 3D-aware semantics. This can ensure a smooth semantic transition during the interpolation. We utilize DDIM inversion to transform their multi-view images into latent noise, extracting geometry and texture features. Then, we interpolate the LoRA parameters and the latent noise to ensure a smooth transition in semantics and features. Meanwhile, we propose a 3D-aware rescheduling strategy during the sampling process to enhance the smooth transition of the interpolated 3D objects. Finally, we utilize 4DGS to reconstruct the interpolation process. The 4DGS can be decomposed into 3DGS to obtain individual 3D objects under different interpolation coefficients. Extensive experiments show that 3DMorphing can effectively interpolate between two 3D objects to generate semantically reasonable interpolated 3D objects with smooth transitions in both geometry and texture.