Native and Compact Structured Latents for 3D Generation: O-Voxel Representation and 4B Parameter Flow Matching Model Achieve Leap in 3D Generation Quality
In December 2025, the research proposed O-Voxel (Omnidirectional Voxel) representation, which can encode geometry and appearance (including PBR materials) of arbitrary topology (open, non-manifold, closed surfaces). Based on this, a sparse compression VAE was designed to achieve high spatial compression ratio and compact latent space. A 4B parameter flow matching model was trained, achieving generation quality far exceeding existing models on public 3D asset datasets, with efficient inference.
This research addresses the bottleneck of representing complex topology and fine appearance in 3D generation through the innovative O-Voxel representation and sparse compression VAE. The 4B parameter flow matching model significantly surpasses existing methods in generation quality while maintaining efficient inference, marking the entry of 3D generation into a high-fidelity, practical stage. This breakthrough will accelerate applications in 3D content creation, gaming, film, and digital twins.
O-Voxel is a sparse voxel structure where each voxel encodes geometry (occupancy probability) and appearance (e.g., color, roughness, metalness and other PBR parameters), capable of representing arbitrary topology (including open surfaces and non-manifolds). The sparse compression VAE based on O-Voxel achieves high compression ratios (e.g., 256^3 voxels compressed to 32^3 latent variables) via sparse convolutions and attention mechanisms while preserving details. The flow matching model uses 4B parameters and is trained on multiple public 3D datasets (e.g., Objaverse, ShapeNet), achieving generation quality significantly superior to existing methods (e.g., GET3D, Point-E) in geometric accuracy and material realism. During inference, the model directly generates O-Voxel latents from noise, then decodes them into explicit meshes or neural fields with high efficiency. The paper does not provide a complete quantitative comparison table but claims to "far exceed existing models."
This technology will directly empower the 3D content creation industry: game developers can quickly generate high-quality 3D assets, film production can reduce special effects costs, and e-commerce can automatically generate 3D product displays. O-Voxel's PBR material support makes it compatible with existing rendering pipelines, lowering deployment barriers. Additionally, the high compression ratio of the sparse compression VAE facilitates deployment on edge devices, promoting AR/VR applications.
It is recommended that 3D content platforms (e.g., game engines, e-commerce) evaluate the integration difficulty of O-Voxel representation with existing pipelines and consider procuring or co-developing generation tools based on this technology. For hardware manufacturers, the sparse computation characteristics of this model may create demand for new acceleration chips. Investment institutions can pay attention to technology startups in this direction.
Attention should be paid to the model's generation capability in complex scenes (e.g., multiple people, dynamics) and the scalability of O-Voxel representation for large-scale scenes. The training and inference costs of the 4B parameter model need further optimization. Open-sourcing model weights and training code will determine community adoption speed. In terms of safety, measures are needed to prevent generation of infringing or inappropriate content.