Diffusion-Based Signed Distance Fields for 3D Shape Generation

Figure 1. Visualization of 3D shape generation for various categories using SDF-Diffusion. SDF-diffusion gradually removes the Gaussian noise and generates high-resolution 3D shapes in the form of an SDF voxel by a two-stage framework. Unlike previous methods, which use point clouds, we can generate 3D meshes without concern of complex post-processing. For visualization purposes, we show less-noise data instead of the initial Gaussian noise.