Spherical Geometry Diffusion: Generating High-quality 3D Face Geometry via Sphere-anchored Representations

TL;DR

This work tackles the difficulty of producing high-quality, topologically sound 3D face geometry in text-driven generation. It introduces a sphere-anchored Spherical Geometry Representation (SGR) that maps 3D facial signals onto a canonical sphere and unwraps them into a structured 2D map, enabling robust connectivity and leveraging mature 2D diffusion models. A two-stage Spherical Geometry Diffusion framework then generates geometry from text and performs geometry-aware texture synthesis conditioned on the geometry, with topology-preserving training via Center-symmetric Padding and Geometric Regularization. The approach achieves state-of-the-art geometric fidelity, efficient inference, and versatile editing capabilities across text-to-geometry, reconstruction, and texture synthesis, demonstrated on multiple 3D face datasets. The method highlights a scalable paradigm for high-fidelity 3D content by regularizing complex 3D structures onto simple manifolds and exploiting 2D generative models.

Abstract

A fundamental challenge in text-to-3D face generation is achieving high-quality geometry. The core difficulty lies in the arbitrary and intricate distribution of vertices in 3D space, making it challenging for existing models to establish clean connectivity and resulting in suboptimal geometry. To address this, our core insight is to simplify the underlying geometric structure by constraining the distribution onto a simple and regular manifold, a topological sphere. Building on this, we first propose the Spherical Geometry Representation, a novel face representation that anchors geometric signals to uniform spherical coordinates. This guarantees a regular point distribution, from which the mesh connectivity can be robustly reconstructed. Critically, this canonical sphere can be seamlessly unwrapped into a 2D map, creating a perfect synergy with powerful 2D generative models. We then introduce Spherical Geometry Diffusion, a conditional diffusion framework built upon this 2D map. It enables diverse and controllable generation by jointly modeling geometry and texture, where the geometry explicitly conditions the texture synthesis process. Our method's effectiveness is demonstrated through its success in a wide range of tasks: text-to-3D generation, face reconstruction, and text-based 3D editing. Extensive experiments show that our approach substantially outperforms existing methods in geometric quality, textual fidelity, and inference efficiency.

Figures (18)

• Figure 1: We present Spherical Geometry Diffusion as a novel framework for 3D faces generation using Spherical Geometry Representation. For Text-based Geometry Generation, we achieve flexible control through text conditions. For Geometry-aware Texture Generation, we create texture conditioning on the geometry through SGR’s alignment.
• Figure 2: Qualitative comparison of reconstructed mesh quality, where the mesh is reconstructed from the implicit function obtained by ImFace++. A magnified view of the marked region is shown on the right.
• Figure 3: Overview. Our pipeline comprises two stages. (a) In the first stage, we construct and compress the Spherical Geometry Representation into Geometric Latent Representations. Geometric Regularization and Center-symmetric Padding are introduced to enhance geometric quality and accelerate convergence. (b) Using these compact latent space, we train a conditional diffusion model through two sequential phases: (1) Text-based Geometry Generation, which generates 3D faces from text prompts, and (2) Geometry-aware Texture Generation, which creates textures conditioned on both text and the geometry SGR. The final latent codes are then decoded to reconstruct 3D meshes of high-quality.
• Figure 4: ($\mathcal{S}$) We obtain the spherical domain $\mathcal{S}$ via Spherical Parameterization $\mathcal{M} \rightarrow \mathcal{S}$, where ($\mathcal{U}$) denotes the unfold SGR grid. Identical colors indicate corresponding regions. ($\mathcal{U}$) The proposed Center-symmetric Padding mirrors the values symmetrically about the center of each edge. The purple color indicates averaging, with identical values marked by the same letters and colors.
• Figure 5: Impact of Center-symmetric Padding.