Source-linked AI summary
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
High-fidelity 3D generation faces data-quality and scalability challenges. UltraShape 1.0 combines data processing with coarse-to-fine diffusion, refining coarse geometry through spatially anchored voxel queries. It reports competitive data-processing and geometry-generation quality against open-source methods, including quality comparable to commercial systems despite limited resources.
Problem
High-fidelity 3D generation is challenged by limited, unevenly represented data and the difficulty of producing fine geometry at higher resolutions.
Source: §1 · p. 2Method
UltraShape 1.0 combines watertight data processing and filtering with two-stage diffusion that refines coarse geometry using spatially anchored voxel queries.
Source: §1 · p. 3, Data Watertightening · p. 3, §2.2 · p. 5Results
UltraShape 1.0 performs competitively with open-source methods in data-processing quality and geometry generation, with quality comparable to commercial systems despite limited computational resources.
Source: §3.3 · p. 11, §4 · p. 12Takeaways & Limitations
The authors present the framework as a step toward more scalable and practical 3D generation systems.
Source: §4 · p. 12
Abstract
from arXiv · showhide
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.
1 Introduction
High-fidelity 3D generation is constrained by scarce, unevenly represented data and the memory and compute costs of sparse representations. UltraShape 1.0 pairs data curation with a coarse-to-fine diffusion pipeline to address data quality and scalable geometry generation.
1 Introduction
- High-quality 3D generation remains difficult because data are scarce and unevenly represented, while sparse representations make memory and computation scale cubically with resolution.These constraints limit fine-grained geometry and robustness at higher resolutions.Source: §1 · p. 2
- UDF-, visibility-, and flood-fill-based remeshing each have limitations, including erroneous component removal, noise, or leakage on unsuitable inputs.The cited approaches use UDFs with an offset [2] [28], ray casting [12] [13] [28], and exterior-seeded filling such as ManifoldPlus [7].Source: §1 · p. 2
- Vector set methods use compact global representations but can miss local detail, whereas sparse voxel methods model local geometry at higher computational and memory cost.Sparse voxel approaches often use two-stage pipelines; Ultra3D [3] refines vector set geometry with sparse voxels, while LATTICE [10] uses structured voxel queries.Source: §1 · pp. 2–3
- UltraShape 1.0 combines watertight geometry processing and data filtering with coarse-to-fine generation using voxel-conditioned refinement.The refinement stage uses structured voxel queries and explicit spatial encoding to separate localization from geometric detail synthesis.Source: §1 · p. 3
2 Method
The method combines voxel-based watertight processing and data curation with a two-stage geometry generator. A vector-set stage captures global structure, and voxel-query refinement focuses on local detail at fixed spatial locations.
Data Watertightening.
- The watertight-processing method resolves topological ambiguities in a sparse voxel domain before surface extraction.Its components include scalable CUDA-parallel reconstruction up to resolutions of 2048^3, automatic hole closing, and thickening open surfaces before signed distance field reconstruction.Source: Data Watertightening · p. 3
- The curation pipeline addresses poor geometry, inconsistent poses, and intricate internal structures through filtering, pose normalization, reconstruction checks, and manual inspection.It reduces an initial pool of about 800K models to approximately 330K valid samples, including 120K identified as high-quality.Source: Data Watertightening · pp. 3–4
2.2 Geometry Generation
- The first generation stage uses a vector-set DiT to represent overall object structure for downstream refinement.Its goal is a coarse but semantically meaningful shape prior, rather than fine-grained detail.Source: §2.2 · pp. 4–5
- The second stage refines geometry on fixed-grid voxel queries, using coarse geometry as spatial anchors and RoPE to encode their coordinates.This separates spatial localization from detail synthesis and reduces the diffusion model's solution space.Source: §2.2 · pp. 5–6
- Voxel-conditioned refinement also supports training-free stylization by applying different image conditions to the two generation stages.The coarse representation lets the second stage sculpt finer stylized details without conflicts between the conditions.Source: §2.2 · pp. 4, 6
3 Experiments
The experiments describe the implementation and data-watertightening pipeline, then evaluate shape generation against open-source and commercial methods and test-time token scaling. UltraShape is reported to deliver high-quality geometry while scaling to larger token counts at inference.
3.1 Implementation Details
- Training uses curvature-aware surface and supervision sampling, rendered image conditioning, 120K filtered Objaverse samples, and progressive token and image-resolution stages.Experiments use eight NVIDIA H20 GPUs; inference uses 32,768 tokens unless otherwise specified.Source: §3.1 · pp. 6–7
3.2 Data Watertightening
- UltraShape consistently closes open surfaces while improving geometric quality, preserving fine details, and avoiding noisy artifacts.Existing approaches can fail to fill large openings or produce noisy, irregular geometry, especially in thin or self-occluded regions.Source: Figure 5 · p. 7, §3.2 · p. 7
3.3 Shape Generation
- Against open-source baselines, UltraShape produces rich details and sharp structural features, with superior geometric fidelity and improved alignment to input images.Qualitative comparisons are shown in Figures 6 and 7.Source: Figure 6 · p. 8, Figure 7 · p. 9, §3.3 · p. 11
- Compared with commercial systems, UltraShape achieves comparable geometric detail and image–geometry consistency despite using public data and limited computational resources.The qualitative comparison is shown in Figure 8.Source: Figure 8 · p. 10, §3.3 · p. 11
- At inference, the DiT generalizes from a relatively small training token budget to larger token counts, producing more detailed geometry and higher-quality surfaces.The reported test-time scaling comparison is in Figure 10.Source: §3.3 · p. 11
- VAE reconstruction quality consistently improves as the number of latent tokens increases.The test-time reconstruction scaling results are shown in Figure 9.Source: Figure 9 · p. 11, §3.3 · p. 11
4 Conclusion
UltraShape combines data processing with a two-stage generation pipeline that separates global structure modeling from local-detail refinement. The report concludes that it achieves strong performance against open-source methods in data processing and geometry generation, with low-cost training resources.
4 Conclusion
- The two-stage pipeline models global structure with a diffusion-based vector-set representation, then refines geometry with voxels for greater detail and scalability.Spatially localized voxel queries and positional encoding condition refinement on coarse geometry.Source: §4 · pp. 11–12
- A watertight processing method and high-quality data filtering improve training-data reliability while preserving geometric details.Source: §4 · p. 11
- UltraShape performs strongly against open-source methods in both data-processing quality and geometry generation, with low-cost training resources.The report describes this as a step toward more scalable and practical 3D generation systems.Source: §4 · p. 12