Skip to content

Source paper

UltraShape 1.0: High-Fidelity 3D Shape Generation via Scalable Geometric Refinement

Tanghui Jia, Dongyu Yan, Dehao Hao and 10 more

Tanghui Jia, Dongyu Yan, Dehao Hao, Yang Li, Kaiyi Zhang, Xianyi He, Lanjiong Li, Jinnan Chen, Lutao Jiang, Qishen Yin, Long Quan, Ying-Cong Chen, Li Yuan

arXiv 2512.21185 · 202525 upvotes on Hugging Face

Research ideas from this paper (3)

Generated by AI. The paper’s authors did not write these.

AI summary

This paper presents UltraShape 1.0, a scalable 3D diffusion framework designed to generate high-fidelity 3D geometry from single images. The authors introduce a two-stage pipeline where a coarse global structure is first synthesized using a vector set representation, followed by a voxel-based refinement stage that decouples spatial localization from detail synthesis to produce intricate, watertight meshes.

How it works (AI summary)

The method employs a robust data curation pipeline involving voxel-based watertightening and VLM-based filtering to create a high-quality dataset. For generation, it uses a cascaded Diffusion Transformer (DiT) architecture: Stage 1 generates a coarse shape using a vector-set DiT (based on Hunyuan3D-2.1), and Stage 2 refines this shape using a voxel-conditioned DiT with Rotary Positional Embeddings (RoPE) to synthesize high-frequency geometric details on a fixed grid.