Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

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

TL;DR

UltraShape 1.0 tackles the bottlenecks of high-fidelity 3D generation by combining thorough data curation with a two-stage diffusion-based geometry pipeline. The two stages separate global-structure modeling from local-detail synthesis using RoPE-encoded, fixed-location voxel queries, enabling scalable, high-fidelity refinements. The data-processing pipeline produces watertight meshes and high-quality samples from public datasets, while the refinement architecture achieves detailed geometry with reduced training resource requirements. Across extensive experiments, UltraShape 1.0 is competitive with open-source baselines and comparable to commercial systems, demonstrating practical scalability for large-scale 3D production.

Abstract

In this report, we introduce UltraShape 1.0, a scalable 3D diffusion framework for high-fidelity 3D geometry generation. The proposed approach adopts a two-stage generation pipeline: a coarse global structure is first synthesized and then refined to produce detailed, high-quality geometry. To support reliable 3D generation, we develop a comprehensive data processing pipeline that includes a novel watertight processing method and high-quality data filtering. This pipeline improves the geometric quality of publicly available 3D datasets by removing low-quality samples, filling holes, and thickening thin structures, while preserving fine-grained geometric details. To enable fine-grained geometry refinement, we decouple spatial localization from geometric detail synthesis in the diffusion process. We achieve this by performing voxel-based refinement at fixed spatial locations, where voxel queries derived from coarse geometry provide explicit positional anchors encoded via RoPE, allowing the diffusion model to focus on synthesizing local geometric details within a reduced, structured solution space. Our model is trained exclusively on publicly available 3D datasets, achieving strong geometric quality despite limited training resources. Extensive evaluations demonstrate that UltraShape 1.0 performs competitively with existing open-source methods in both data processing quality and geometry generation. All code and trained models will be released to support future research.

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

TL;DR

UltraShape 1.0 tackles the bottlenecks of high-fidelity 3D generation by combining thorough data curation with a two-stage diffusion-based geometry pipeline. The two stages separate global-structure modeling from local-detail synthesis using RoPE-encoded, fixed-location voxel queries, enabling scalable, high-fidelity refinements. The data-processing pipeline produces watertight meshes and high-quality samples from public datasets, while the refinement architecture achieves detailed geometry with reduced training resource requirements. Across extensive experiments, UltraShape 1.0 is competitive with open-source baselines and comparable to commercial systems, demonstrating practical scalability for large-scale 3D production.

Abstract

In this report, we introduce UltraShape 1.0, a scalable 3D diffusion framework for high-fidelity 3D geometry generation. The proposed approach adopts a two-stage generation pipeline: a coarse global structure is first synthesized and then refined to produce detailed, high-quality geometry. To support reliable 3D generation, we develop a comprehensive data processing pipeline that includes a novel watertight processing method and high-quality data filtering. This pipeline improves the geometric quality of publicly available 3D datasets by removing low-quality samples, filling holes, and thickening thin structures, while preserving fine-grained geometric details. To enable fine-grained geometry refinement, we decouple spatial localization from geometric detail synthesis in the diffusion process. We achieve this by performing voxel-based refinement at fixed spatial locations, where voxel queries derived from coarse geometry provide explicit positional anchors encoded via RoPE, allowing the diffusion model to focus on synthesizing local geometric details within a reduced, structured solution space. Our model is trained exclusively on publicly available 3D datasets, achieving strong geometric quality despite limited training resources. Extensive evaluations demonstrate that UltraShape 1.0 performs competitively with existing open-source methods in both data processing quality and geometry generation. All code and trained models will be released to support future research.
Paper Structure (16 sections, 10 figures)

This paper contains 16 sections, 10 figures.

Table of Contents

  1. Introduction
  2. Method
  3. Data Curation Pipeline
  4. Data Watertightening.
  5. Data Filtering.
  6. Geometry Generation
  7. Coarse Structure Generation.
  8. Geometry Refinement.
  9. Training-Free Stylization.
  10. Experiments
  11. Implementation Details
  12. Data Preparation.
  13. Detailed Settings
  14. Data Watertightening
  15. Shape Generation
  16. ...and 1 more sections

Figures (10)

  • Figure 1: High-quality 3D assets generated by UltraShape 1.0. Best viewed with zoom-in.
  • Figure 2: Overview of UltraShape 1.0 pipeline, where Enc. and Dec. represent the encoder and decoder of our VAE. The superscript “2” on the upper right corner denotes the Stage-2 model. MC means marching cube.
  • Figure 3: Comparison of our refined mesh against the coarse mesh generated from the first stage. Best viewed with zoom-in.
  • Figure 4: Stylized results generated using our training-free stylization method. The refined and stylized meshes in the last three columns are generated by refining the coarse mesh in the first column using image conditions on the bottom-right corner. Best viewed with zoom-in.
  • Figure 5: Comparison of our watertightening method against other approaches. Best viewed with zoom-in.
  • ...and 5 more figures