
Explore TRELLIS, Microsoft Research's groundbreaking AI model that generates high-quality 3D assets from text or images in seconds.
Creating 3D models used to require hours of manual work in specialized software. TRELLIS 3D changes everything — it's an AI model developed by Microsoft Research that generates production-ready 3D assets from a single image or text prompt in just a few seconds.
Whether you're a game developer, 3D printing enthusiast, or digital artist, TRELLIS offers a new way to create 3D content that was previously impossible without years of 3D modeling expertise.
TRELLIS stands for Structured LATent — a reference to its core technical innovation: a structured latent representation that captures both the geometry and appearance of 3D objects in a compact, flexible format.
This representation is what makes TRELLIS uniquely powerful compared to other 3D generation approaches.
TRELLIS generates 3D assets through a two-stage process:
3D technology specialists focused on AI-powered 3D model generation, format conversion, and browser-based 3D rendering. We test and review 3D tools so you don't have to.
Join the community
Subscribe to our newsletter for the latest news and updates
A Rectified Flow Transformer predicts which parts of 3D space are occupied — essentially sketching the rough shape of the object. This determines where the object exists in three-dimensional space.
For each occupied region, the model generates a structured latent representation that captures fine geometric details and surface textures. This is where the visual richness comes from.
One of TRELLIS's most distinctive features is that the same generation can be decoded into multiple output formats:
| Format | Best For |
|---|---|
| 3D Gaussian Splatting | Real-time rendering, web viewers |
| Radiance Fields (NeRF) | Photorealistic visualization |
| Mesh (GLB) | Game engines, 3D printing, AR/VR |
This means you don't need to choose your output format upfront — generate once, export as needed.
Microsoft released two major versions of TRELLIS:
| Feature | TRELLIS v1 | TRELLIS 2 |
|---|---|---|
| Parameters | ~2 billion | 4 billion |
| Generation speed | ~10 seconds | ~3 seconds |
| Max resolution | Standard | Up to 1536³ voxels |
| Material quality | Standard | High-fidelity PBR |
| Open source license | Custom | MIT |
| Release date | December 2024 | December 2025 |
TRELLIS 2 introduces O-Voxel (Omni-Voxel) representation and a Sparse Compression VAE that achieves 16x spatial compression, enabling higher quality with faster generation.
Describe what you want in natural language and get a 3D model. From "a medieval sword with a ruby pommel" to "a cute robot holding a balloon" — if you can describe it, TRELLIS can generate it.
Upload a single photo and TRELLIS reconstructs a full 3D model. This works for objects, characters, products, and more. The model infers depth, occluded surfaces, and texture details from a single viewpoint.
Unlike most 3D generation models, TRELLIS supports local editing — you can modify specific parts of a generated 3D asset without regenerating the entire model. This is invaluable for iterative design workflows.
Create characters, props, and environment assets in seconds instead of hours. Export as GLB files that work directly in Unity, Unreal Engine, and web-based game frameworks.
TRELLIS generates watertight geometry and manifold meshes that can be exported as STL files for 3D printing — no manual cleanup required.
Generate 3D product models from catalog photos. Let customers interact with products in 3D on your website, increasing engagement and reducing return rates.
Create immersive 3D assets for augmented and virtual reality applications. The GLB output format is ideal for WebXR and other AR/VR platforms.
TRELLIS stands out in the 3D generation landscape for several reasons:
The model was trained on 500,000 diverse 3D objects and has been recognized at CVPR 2025, one of the top computer vision conferences in the world.
Learn how to get the most out of TRELLIS 2 with our step-by-step guides:
TRELLIS 2 is fully open source under the MIT license. You can find the code on GitHub:
Pre-trained model weights are available on Hugging Face.
You can also try TRELLIS without setting up anything locally. Our platform provides an easy-to-use interface for generating 3D models from text or images — no coding required.
For those interested in the research behind TRELLIS:
The research was led by the Spatial Intelligence Group at Microsoft Research Asia (MSRA), with key contributors including Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, and Jiaolong Yang.
Ready to create your first 3D model? You don't need to install anything or write a single line of code.
Our online TRELLIS 3D generator lets you:
| Feature | Self-Hosted | Our Platform |
|---|---|---|
| Setup time | 2-4 hours | 0 minutes |
| GPU required | Yes (24GB+ VRAM) | No |
| Technical knowledge | Python, CUDA | None |
| Cost | Hardware + electricity | Free tier available |
| Speed | Depends on your GPU | Optimized for speed |
Don't have a high-end GPU? No problem. Our cloud infrastructure runs TRELLIS 2 on optimized hardware, delivering 3-second generations without any setup.
TRELLIS 3D represents a fundamental shift in how 3D content is created. What once required specialized skills and expensive software can now be done in seconds with AI. Whether you're building games, creating AR experiences, or prototyping products, TRELLIS makes 3D creation accessible to everyone.
As the technology continues to evolve with TRELLIS 2 and beyond, we can expect even faster generation, higher quality, and new creative possibilities that we haven't yet imagined.