Title: SymTRELLIS: Enforcing Symmetry via Voxel Latents for 3D Generation

Abstract:

While single-view 3D generative models have delivered remarkable visual fidelity, they frequently fail to meet structural or functional criteria, often resulting in physically impractical outputs. A critical oversight is symmetry; even minor deviations can render a generated model unusable for practical applications. To address this, we introduce SymTRELLIS, a technique that imposes arbitrary finite point group symmetries—including rotational, reflectional, and polyhedral constraints—during the flow-based 3D generation process of TRELLIS.2. Notably, this is achieved without the need to retrain the foundational VAE or flow models.

The core innovation of SymTRELLIS involves approximating the latent-space effects of spatial transformations as a learned linear operator applied to voxel latents. This is realized through a lightweight spatial-transform latent mapper, which is trained on standard, non-symmetric 3D datasets. During the generation phase, symmetry is guaranteed through a process termed "velocity symmetrization," where predicted flow velocities are averaged across all symmetry-equivalent transformations at every step of the Ordinary Differential Equation (ODE) integration.

Users can provide symmetry specifications manually, or the system can estimate them automatically from an initial TRELLIS.2 output. This flexibility allows for intentional manipulation of folds and structures that may not be evident in the source image. In evaluations on a curated benchmark comprising 266 strictly symmetric objects—featuring 2- to 20-fold rotations and various polyhedral symmetry groups—SymTRELLIS significantly outperformed TRELLIS.2, Hunyuan3D-2.1, and TripoSG across all symmetry error metrics. Crucially, these improvements in symmetry adherence were achieved while maintaining reconstruction accuracy levels comparable to the original base model.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC