UDFStudio: A Unified Framework of Datasets, Benchmarks and Generative Models for Unsigned Distance Functions

Area of Science:

• Computer Graphics
• Artificial Intelligence
• 3D Geometry Processing

Background:

• Unsigned distance functions (UDFs) are effective for open surfaces but lack dedicated 3D generative models.
• Existing 3D datasets primarily feature watertight meshes, hindering open-surface research.
• Generating diverse open-surface 3D content is limited by current methodologies.

Purpose of the Study:

• Introduce UDiFF, a novel diffusion-based 3D generative model for UDFs.
• Enable conditional and unconditional generation of textured 3D shapes with open surfaces.
• Establish standardized benchmarks for UDF-based open-surface generation and reconstruction.

Main Methods:

• Developed UDiFF, a diffusion model operating in the spatial-frequency domain.
• Implemented a learnable wavelet transform for UDF generation, optimized from data.
• Created the UWings dataset with 1,509 winged creature 3D models.

Main Results:

• UDiFF successfully generates textured 3D shapes with open surfaces.
• The learnable wavelet transform outperforms manual selection in UDF generation.
• The UWings dataset provides a benchmark for evaluating UDF-based methods.

Conclusions:

• UDiFF advances 3D generative modeling for open surfaces using UDFs.
• The data-driven wavelet approach enhances UDF representation learning.
• The UWings dataset and benchmarks facilitate future research in open-surface reconstruction.