🚀 Microsoft Open-Source • 4B Parameters • MIT License

Transform Any Image to 3D — Powered by Microsoft AI

TRELLIS-2 generates professional high-fidelity 3D models with PBR materials in seconds. The most advanced open-source image-to-3D AI model with 4 billion parameters.

3-Second Generation at 512³
🎨 Full PBR Material Support
📦 GLB/OBJ/PLY Export
🔓 100% Free & Open Source
4B
Parameters
15K+
GitHub Stars
1536³
Max Resolution
~3s
Generation Time
TRELLIS-2 Live Demo — trellis-2.org
Trusted by 50,000+ Creators Worldwide
15,000+ GitHub Stars
📥
2.3M+ Monthly Downloads
🏢
Microsoft Research Backed
🏆
CVPR'25 Spotlight Paper

What is TRELLIS-2?

TRELLIS-2 is Microsoft Research's revolutionary 4-billion parameter open-source AI model that transforms single images into high-fidelity 3D assets. Built on the innovative O-Voxel (Omni-Voxel) architecture, it represents the cutting edge of 3D generative AI technology.

Unlike traditional 3D modeling tools that require hours of manual work, TRELLIS-2 leverages advanced Flow-Matching Transformers and Structured 3D Latent representations (SLAT) to generate production-ready 3D models with full PBR materials in seconds.

Key Innovation: TRELLIS-2's O-Voxel representation enables handling of arbitrary topologies including open surfaces, non-manifold geometry, and enclosed interiors — capabilities that surpass traditional iso-surface methods.

Explore Features →
// TRELLIS-2 Pipeline
🖼️
Image Input
🧠
AI Processing
📦
3D Output

TRELLIS-2 Key Features

Everything you need to create stunning 3D assets with the most advanced open-source AI model available.

🚀

4B Parameter Model

The largest open-source image-to-3D model with 4 billion parameters, delivering unmatched quality and detail in generated 3D assets.

Lightning Fast Generation

Generate 512³ resolution models in ~3 seconds on H100 GPU. Full 1536³ high-resolution output in approximately 60 seconds.

🎨

Full PBR Material Support

Complete physically-based rendering materials including Base Color, Roughness, Metallic, and Opacity for realistic rendering.

🔮

O-Voxel Architecture

Revolutionary Omni-Voxel representation that handles arbitrary topologies, open surfaces, and complex geometries.

📤

Multi-Format Export

Export to GLB, OBJ, PLY, Radiance Fields, and 3D Gaussians. Direct integration with Blender, Unity, and Unreal Engine.

🔓

MIT Open Source License

Completely free and open-source under MIT license. Use for personal projects, commercial applications, and research.

Supported Output Formats

TRELLIS-2 supports multiple output formats for seamless integration with your existing workflow.

Format Type PBR Support Max Resolution Best For
GLB Mesh ✓ Full 4096×4096 textures Game engines, Web3D, AR/VR
OBJ Mesh ✓ MTL Unlimited 3D modeling software
PLY Point Cloud ✓ Vertex Unlimited Scientific visualization
Radiance Fields Neural ✓ Implicit Continuous Novel view synthesis
3D Gaussians Gaussian Splatting ✓ Native Adaptive Real-time rendering

Blazing Fast Performance

TRELLIS-2 delivers industry-leading generation speeds while maintaining exceptional quality. Optimized for NVIDIA GPUs with support for various VRAM configurations.

~3s
512³ Resolution
~17s
1024³ Resolution
~60s
1536³ Resolution
16×
Latent Compression
// Generation Speed Comparison (H100 GPU)
TRELLIS-2
3s
Meshy
45s
Tripo AI
60s
TripoSR
8s

TRELLIS-2 Use Cases

From game development to e-commerce, TRELLIS-2 empowers creators across industries.

🎮

Game Development

Rapidly prototype game assets, characters, and environments. Export directly to Unity and Unreal Engine.

🥽

AR/VR Content

Create immersive 3D experiences for augmented and virtual reality applications with optimized meshes.

🖨️

3D Printing

Generate watertight meshes ready for 3D printing. Perfect for prototyping and manufacturing.

🛒

E-Commerce

Transform product photos into interactive 3D models for online stores and AR product previews.

🏗️

Architecture

Quickly generate 3D visualizations of buildings, interiors, and landscape designs from sketches.

🎬

Film & Animation

Accelerate pre-visualization and asset creation for film and animation production pipelines.

📚

Education

Create educational 3D models for interactive learning experiences and virtual classrooms.

🔬

Research

Advance 3D AI research with access to state-of-the-art model architecture and training techniques.

TRELLIS-2 vs Competitors

See how TRELLIS-2 compares to other leading 3D AI generation tools in the market.

Feature TRELLIS-2 Meshy Tripo AI CSM.ai TripoSR
Parameters 4B Unknown 20B Unknown ~500M
Open Source
Price Free $20/mo $15.9/mo $20/mo Free
PBR Materials ✓ Full Limited
Max Resolution 1536³ 1024³ 1024³ 512³ 512³
Generation Speed ~3s ~45s ~60s ~30s ~8s
Local Deployment
Transparency Support

Technical Specifications

Model Architecture

  • Model Size4 Billion Parameters
  • ArchitectureFlow-Matching Transformer
  • 3D RepresentationO-Voxel (Omni-Voxel)
  • Latent SpaceSLAT (Structured 3D Latent)
  • Compression16× Spatial Downsampling
  • Visual EncoderDINOv2 Multi-view

System Requirements

  • Operating SystemLinux (Ubuntu 20.04+)
  • GPUNVIDIA 24GB+ VRAM
  • CUDA Version12.4+
  • Python3.8+
  • Storage50GB+ Free Space
  • RAM32GB+ Recommended

Quick Installation Guide

Get TRELLIS-2 running locally in just a few steps. Follow our streamlined setup process.

1

Clone Repository

Clone the official TRELLIS-2 GitHub repository to your local machine.

git clone https://github.com/microsoft/TRELLIS.2.git && cd TRELLIS.2
2

Setup Environment

Create a conda environment and install the required dependencies.

conda create -n trellis python=3.10 && conda activate trellis && pip install -r requirements.txt
3

Run Generation

Start generating 3D models from your images using the simple Python API.

python run.py --image your_image.png --output output.glb
View Full Documentation →

Developer API & Integration

TRELLIS-2 provides a clean Python API for seamless integration into your existing workflows. Build custom applications, automate asset pipelines, or integrate with cloud services.

  • Simple Python API with type hints
  • Hugging Face Transformers compatibility
  • NVIDIA NIM integration for 20% faster inference
  • Batch processing support
  • Customizable generation parameters
  • Webhook callbacks for async workflows
Hugging Face Model →
Python
from trellis import TrellisImageTo3D

# Initialize TRELLIS-2 pipeline
pipeline = TrellisImageTo3D.from_pretrained(
    "microsoft/TRELLIS-image-large"
)

# Generate 3D from image
result = pipeline.generate(
    image="input.png",
    resolution="1024",
    output_format="glb"
)

# Export with PBR materials
result.export("output.glb")
print("3D model generated!")

O-Voxel Architecture Explained

Understanding TRELLIS-2's innovative Omni-Voxel representation and Flow-Matching Transformer architecture.

🖼️

Image Encoder

DINOv2 extracts multi-view visual features

🧊

Sparse 3D VAE

Compresses to structured latent space

🔄

Flow Transformer

Generates O-Voxel representation

📦

Mesh Decoder

Outputs textured 3D mesh

SLAT Representation

Structured 3D Latent representation combines geometry and appearance in a unified sparse voxel grid, enabling efficient compression and high-fidelity reconstruction.

Flow Matching

Rectified Flow Transformers provide faster and more stable training compared to traditional diffusion models, resulting in higher quality outputs.

O-Voxel Benefits

Handles arbitrary topologies including open surfaces, non-manifold geometry, transparency, and enclosed interiors that traditional methods cannot represent.

Seamless Workflow Integration

TRELLIS-2 integrates perfectly with your favorite 3D tools and game engines.

🔶

Blender

Direct GLB import with full PBR material support

🎮

Unity

Import ready-to-use game assets with optimized meshes

⚙️

Unreal Engine

Production-quality assets for UE5 projects

🌐

Three.js

Web-ready 3D models for browser applications

🤗

Hugging Face

Cloud inference via Spaces and Inference API

🟢

NVIDIA NIM

Optimized deployment with 20% faster inference

🐳

Docker

Containerized deployment for production environments

☁️

Cloud APIs

fal.ai, Replicate, and custom cloud deployments

What Developers Say

Hear from developers and artists using TRELLIS-2 in production.

TRELLIS-2 completely changed our asset pipeline. We can now prototype 3D game assets in seconds instead of hours. The quality is incredible for an open-source model.
JD
James Davidson Indie Game Developer
The PBR material support and mesh quality are phenomenal. Finally an open-source solution that rivals commercial offerings. Microsoft really delivered with TRELLIS-2.
SK
Sarah Kim 3D Artist & Tech Artist
As a researcher, I appreciate the open architecture and reproducible results. The O-Voxel representation is a genuine innovation in 3D generation AI.
ML
Dr. Michael Liu AI Researcher

Frequently Asked Questions

Everything you need to know about TRELLIS-2 AI 3D generation.

What is TRELLIS-2?
+

TRELLIS-2 is Microsoft Research's 4-billion parameter open-source AI model for generating high-fidelity 3D assets from single images. It uses innovative O-Voxel architecture and Flow-Matching Transformers to create production-ready 3D models with PBR materials in seconds.

Is TRELLIS-2 free to use?
+

Yes, TRELLIS-2 is completely free and open-source under the MIT license. You can use it for personal projects, commercial applications, and research without any cost or subscription fees.

What GPU do I need for TRELLIS-2?
+

TRELLIS-2 requires an NVIDIA GPU with at least 24GB VRAM (such as RTX 4090, A100, or H100). For optimal performance at 1536³ resolution, an H100 GPU is recommended.

What output formats are supported?
+

TRELLIS-2 supports multiple output formats including GLB (with PBR textures up to 4096×4096), OBJ, PLY, Radiance Fields, and 3D Gaussians. Models are compatible with Blender, Unity, Unreal Engine, and other 3D software.

How fast is TRELLIS-2?
+

TRELLIS-2 is incredibly fast. On an H100 GPU, it generates 512³ resolution models in approximately 3 seconds, 1024³ in about 17 seconds, and 1536³ in around 60 seconds.

Can I use TRELLIS-2 commercially?
+

Yes, TRELLIS-2 is released under the MIT license which permits commercial use. However, some NVIDIA dependencies may have separate restrictions, so check the specific components you use.

What is O-Voxel?
+

O-Voxel (Omni-Voxel) is TRELLIS-2's innovative sparse voxel representation that encodes both geometry and appearance in a unified structure. It enables handling of arbitrary topologies including open surfaces, non-manifold geometry, and enclosed interiors.

How does TRELLIS-2 compare to Meshy?
+

TRELLIS-2 offers several advantages: it's completely free and open-source, has 4B parameters, supports higher resolution output (1536³ vs 1024³), faster generation (~3s vs ~45s), and allows local deployment. Unlike Meshy's $20/month subscription, TRELLIS-2 has no credit limits or fees.

Learning Resources

Tutorials, documentation, and guides to help you master TRELLIS-2.

📖

Official Documentation

Comprehensive guides and API reference

Read Docs →
🎬

Video Tutorials

Step-by-step video guides on YouTube

Watch Videos →
📄

Research Paper

CVPR'25 Spotlight paper on arXiv

Read Paper →
💬

Community Forum

GitHub Discussions and Reddit

Join Discussion →

Join the TRELLIS-2 Community

Connect with thousands of developers, artists, and researchers using TRELLIS-2 to create amazing 3D content.

15K+
GitHub Stars
50K+
Users
2.3M+
Downloads
// GitHub Activity (Last 12 Weeks)

Simple, Transparent Pricing

TRELLIS-2 is open-source and free. Choose how you want to use it.

Community
Open-source self-hosted
$0 forever
  • Unlimited generations
  • Full 4B model access
  • All output formats
  • Local deployment
  • MIT license
  • Community support
Get Started Free
API Access
Cloud inference services
Varies per provider
  • fal.ai integration
  • NVIDIA NIM API
  • Replicate hosting
  • Production SLA
  • Priority support
  • Custom deployment
View Providers

Latest Updates

Stay informed about TRELLIS-2 releases and improvements.

December 2024
TRELLIS.2 Released

Major update with 4B parameter model, O-Voxel architecture, and full PBR material support.

December 2024
CVPR'25 Spotlight Paper

TRELLIS research accepted as Spotlight paper at CVPR 2025 conference.

December 2024
Hugging Face Integration

Official Hugging Face Space launched with free cloud inference demo.

December 2024
NVIDIA NIM Support

NVIDIA NIM integration with 20% inference speedup on optimized infrastructure.

Development Roadmap

Upcoming features and improvements planned for TRELLIS-2.

Q1 2025
  • Text-to-3D support
  • Multi-image input
  • Improved mesh topology
  • Windows native support
Q2 2025
  • Animation support
  • Auto-rigging
  • LoRA fine-tuning
  • Reduced VRAM mode
Q3 2025
  • Scene generation
  • Physics simulation
  • Material editing
  • Style transfer
Q4 2025
  • Video to 3D
  • Real-time preview
  • Enterprise features
  • Extended API

Common Troubleshooting

Solutions to frequently encountered issues when using TRELLIS-2.

⚠️ CUDA Out of Memory

Running out of GPU memory during generation.

Solution:

Reduce resolution to 512³ or use --low-vram flag. Ensure no other GPU processes are running.

⚠️ Import Error: o-voxel

Missing o-voxel dependency after installation.

Solution:

Run: pip install git+https://github.com/microsoft/o-voxel.git

⚠️ Mesh Has Holes

Generated mesh has small holes or artifacts.

Solution:

Use the included mesh_postprocess.py script to create watertight meshes for 3D printing.

⚠️ Slow Generation

Generation taking longer than expected.

Solution:

Ensure CUDA is properly installed. Use torch.compile() for significant speedup on supported GPUs.

TRELLIS-2 Best Practices

Tips and tricks to get the best results from TRELLIS-2.

🖼️

Use High-Quality Images

Start with clear, well-lit images with the subject centered. Remove busy backgrounds for better results.

📐

Match Resolution to Use Case

Use 512³ for quick previews, 1024³ for game assets, and 1536³ for high-detail renders and 3D printing.

🔄

Post-Process Meshes

Use Blender's decimate modifier to optimize poly count for game engines. Apply remeshing for cleaner topology.

🎨

Optimize PBR Materials

Bake textures to lower resolutions for web and mobile. Use texture atlasing for multiple objects.

Batch Processing

Use the batch API for processing multiple images efficiently. Leverage GPU parallelism when available.

💾

Cache Model Weights

Set HF_HOME environment variable to cache model weights locally and avoid repeated downloads.

Open Source MIT License

TRELLIS-2 is released under the permissive MIT license, giving you complete freedom to use, modify, and distribute the software.

📜
MIT License Free for commercial and personal use
Commercial use allowed
Modification allowed
Distribution allowed
Private use allowed

Powered by Industry Leaders

TRELLIS-2 is developed and supported by leading technology organizations.

About TRELLIS-2.org

TRELLIS-2.org is an independent resource hub dedicated to Microsoft's TRELLIS-2 3D generation AI model. We provide tutorials, documentation, demos, and community resources to help developers and artists leverage this powerful open-source technology.

Our mission is to make 3D AI generation accessible to everyone by providing comprehensive guides, best practices, and a supportive community for users of all skill levels.

Disclaimer: This is an independent fan site and is not officially affiliated with Microsoft Corporation. TRELLIS-2 is developed by Microsoft Research.

Get in Touch

Have questions or feedback? Reach out through these channels.

📧

Email

contact@trellis-2.org

💬

GitHub Issues

Report bugs and feature requests

🎮

Discord

Join our community server

Ready to Create Amazing 3D?

Join thousands of developers and artists using TRELLIS-2 to transform images into stunning 3D assets. Start creating for free today.