UltraShape 1.0

From Peking University Yuan Group

High-Fidelity 3D Shape Generation Framework – Scalable Two-Stage Diffusion for Detailed Geometry from Text or Images

Pricing Model

Starting Price

$0/Month

Visit UltraShape 1.0
Last Updated: January 9, 2026
By Zelili AI

About This AI

UltraShape 1.0 is an advanced open-source 3D diffusion framework developed by the PKU Yuan Group for generating high-fidelity 3D shapes with exceptional geometric quality.

It employs a scalable two-stage pipeline: first synthesizing a coarse global structure, then refining it for fine-grained details using voxel-based refinement with positional anchors via RoPE.

A comprehensive data processing pipeline enhances public 3D datasets through novel watertight processing, hole filling, thin structure thickening, low-quality sample removal, and high-quality filtering while preserving details.

Trained exclusively on public datasets, it achieves strong performance despite limited resources, supporting text-to-3D and image-to-3D generation with outputs in mesh (via marching cubes) or SDF grid formats.

Key strengths include decoupling spatial localization from detail synthesis for efficient refinement, competitive results against other open-source 3D generators, and full release of code, pre-trained models, and inference/training scripts.

Announced via arXiv paper on December 24, 2025 (v1), with inference code/models released December 26, 2025, and training code on December 31, 2025.

Hosted on GitHub (PKU-YuanGroup/UltraShape-1.0) and Hugging Face, it’s ideal for researchers, developers, and creators in 3D AI, game development, AR/VR, and digital content needing precise, watertight, high-detail 3D models without proprietary dependencies.

Key Features

  1. Two-stage diffusion pipeline: Coarse global structure synthesis followed by voxel-based fine-grained refinement
  2. Scalable geometric refinement: Decouples spatial localization from detail synthesis using RoPE-encoded voxel queries
  3. Advanced data processing: Novel watertight processing, hole filling, thin structure thickening, and quality filtering on public 3D datasets
  4. High-fidelity output: Produces detailed, watertight meshes or SDF grids via extended shape VAE and marching cubes
  5. Text-to-3D and image-to-3D support: Generates 3D shapes from text prompts or reference images
  6. Competitive open-source performance: Strong geometric quality rivaling other methods despite public data only
  7. Full code and models release: Inference, training scripts, and pre-trained weights available on GitHub and Hugging Face
  8. Efficient inference: Structured solution space reduces complexity for detailed generation

Price Plans

  1. Free ($0): Fully open-source with code, pre-trained models, inference, and training scripts available on GitHub and Hugging Face; no usage fees

Pros

  1. Exceptional geometric quality: Achieves high-fidelity details and watertight meshes competitive with open-source leaders
  2. Fully open-source: Code, models, and pipeline freely available under permissive license for research and extension
  3. Scalable approach: Two-stage design enables high detail without excessive compute
  4. Robust data handling: Novel watertight and filtering pipeline improves training data quality significantly
  5. Public dataset only: Strong results without proprietary data, promoting accessibility
  6. Active development: Rapid releases of paper, inference, and training code in late December 2025

Cons

  1. Local setup required: Needs GPU and Python environment; no hosted web demo mentioned
  2. Recent release: Limited community adoption, examples, or third-party integrations yet
  3. Compute intensive: Refinement stage and training likely require powerful hardware
  4. No user metrics: As a new academic model, no widespread user numbers or real-world adoption stats
  5. Research-focused: May lack polished UI or easy one-click tools compared to commercial 3D generators
  6. Potential artifacts: Like all diffusion models, complex prompts may yield inconsistencies

Use Cases

  1. 3D content creation: Generate high-detail models for games, AR/VR, animation, or product visualization
  2. Research in 3D AI: Benchmarking, extending diffusion frameworks, or studying geometric refinement
  3. Digital asset prototyping: Quick text/image-to-3D for concept art or design iteration
  4. Academic studies: Training or fine-tuning on improved 3D datasets with watertight processing
  5. Game development: Creating detailed meshes for environments, characters, or props
  6. Industrial design: Prototyping shapes from descriptions or sketches

Target Audience

  1. AI researchers: Studying 3D diffusion, geometry generation, or data processing pipelines
  2. 3D artists and developers: Needing high-fidelity open-source models for assets
  3. Game/AR/VR creators: Generating detailed watertight 3D shapes efficiently
  4. Academic groups: Reproducing or building on PKU Yuan Group's work
  5. Open-source enthusiasts: Experimenting with latest 3D generative frameworks

How To Use

  1. Clone repo: git clone https://github.com/PKU-YuanGroup/UltraShape-1.0.git
  2. Install dependencies: Follow README for Python env setup, likely torch, diffusers, etc.
  3. Download models: Pull pre-trained weights from Hugging Face or links in repo
  4. Run inference: Use provided scripts for text/image-to-3D generation with prompts
  5. Process data: Apply watertight pipeline if preparing custom 3D datasets
  6. Train/fine-tune: Use released training code on public datasets if needed
  7. View results: Export meshes/SDF grids; visualize in tools like Blender

How we rated UltraShape 1.0

  • Performance: 4.6/5
  • Accuracy: 4.7/5
  • Features: 4.5/5
  • Cost-Efficiency: 5.0/5
  • Ease of Use: 4.0/5
  • Customization: 4.8/5
  • Data Privacy: 5.0/5
  • Support: 4.2/5
  • Integration: 4.4/5
  • Overall Score: 4.6/5

UltraShape 1.0 integration with other tools

  1. Hugging Face: Pre-trained models hosted for easy download and inference with diffusers library
  2. GitHub: Full source code repository for cloning, contributing, or extending the framework
  3. Blender/MeshLab: Export meshes/SDF grids for import into popular 3D editing and visualization software
  4. PyTorch Ecosystem: Built on torch/diffusers for seamless integration with other AI 3D pipelines
  5. Research Tools: Compatible with datasets like ShapeNet or Objaverse for training/extensions

Best prompts optimised for UltraShape 1.0

  1. A highly detailed futuristic cyberpunk cityscape building with neon lights and intricate metallic structures, realistic textures, high resolution 3D model
  2. An elegant ancient Chinese dragon statue coiled around a pillar, ornate scales, flowing whiskers, mythical and majestic, watertight mesh ready for rendering
  3. A realistic human hand holding a glowing crystal orb, fine skin details, veins, dynamic lighting, high-fidelity 3D geometry
  4. A sci-fi spaceship interior cockpit with control panels, holographic displays, ergonomic seats, detailed mechanical parts, ultra high quality
  5. A cute cartoon fox character in standing pose, fluffy fur, expressive eyes, vibrant colors, clean topology suitable for animation
UltraShape 1.0 delivers impressive high-fidelity 3D shape generation through its scalable two-stage diffusion and robust watertight data pipeline, achieving strong results on public datasets. Fully open-source with code and models released in late 2025, it’s a valuable resource for researchers and developers in 3D AI. Setup requires technical knowledge, but the geometric quality and refinement approach make it competitive for detailed mesh creation.

FAQs

  • What is UltraShape 1.0?

    UltraShape 1.0 is an open-source scalable 3D diffusion framework from PKU Yuan Group for high-fidelity 3D geometry generation using a two-stage coarse-to-fine pipeline and advanced watertight data processing.

  • When was UltraShape 1.0 released?

    The technical report was published on arXiv December 24, 2025; inference code and models released December 26, 2025; training code December 31, 2025.

  • Is UltraShape 1.0 free and open-source?

    Yes, it’s fully open-source with code, pre-trained models, inference, and training scripts available on GitHub (PKU-YuanGroup/UltraShape-1.0) and Hugging Face at no cost.

  • What does UltraShape 1.0 generate?

    It generates high-detail, watertight 3D meshes or SDF grids from text prompts or images, focusing on fine-grained geometry with strong realism.

  • How do I use UltraShape 1.0?

    Clone the GitHub repo, install dependencies (PyTorch/diffusers), download models, and run inference scripts with text/image inputs; technical setup required.

  • Who developed UltraShape 1.0?

    Developed by researchers from Peking University Yuan Group (PKU-YuanGroup), including Tanghui Jia, Li Yuan, and collaborators.

  • Does UltraShape 1.0 support image-to-3D?

    Yes, it supports both text-to-3D and image-to-3D generation for creating detailed 3D shapes from reference images or prompts.

  • What makes UltraShape 1.0 unique?

    Its two-stage pipeline with voxel refinement using RoPE anchors, plus novel watertight data processing for improved quality on public datasets.

Newly Added Tools​

Qwen-Image-2.0

$0/Month

Qodo AI

$0/Month

Codiga

$10/Month

Tabnine

$59/Month
UltraShape 1.0 Alternatives

Qwen-Image-2.0

$0/Month

GLM-OCR

$0/Month

Lummi AI

$10/Month

About Author

Hi Guys! We are a group of ML Engineers by profession with years of experience exploring and building AI tools, LLMs, and generative technologies. We analyze new tools not just as a user, but as someone who understands their technical depth and real-world value.We know how overwhelming these tools can be for most people, that’s why we break down complex AI concepts into simple, practical insights. Our goal is to help you discover these magical AI tools that actually save your time and make everyday work smarter, not harder.“We don’t just write about AI: We build, test and simplify it for you.”