TRELLIS 3D

What is TRELLIS

TRELLIS is a state-of-the-art native 3D generative model developed through a collaboration between Microsoft Research, Tsinghua University, and the University of Science and Technology of China (USTC). It stems from the research paper Structured 3D Latents for Scalable and Versatile 3D Generation, which was selected as a Highlight paper at CVPR 2025, one of the most prestigious conferences in computer vision and pattern recognition. The official website https://trellis3d.github.io is the only authorized platform for the TRELLIS project, providing access to the full paper, arXiv preprint, code repository, and interactive demo. At its core, TRELLIS is built on two groundbreaking innovations: a unified Structured Latent (SLAT) representation and rectified flow transformers specifically tailored for 3D data. This unique architecture enables it to generate high-quality 3D assets with unprecedented versatility, addressing critical limitations of previous 3D generation methods.

Core Functions of TRELLIS

TRELLIS offers a comprehensive suite of 3D generation and editing capabilities that streamline the entire 3D asset creation workflow. It supports text-to-3D asset generation, allowing users to bring descriptive ideas to life as detailed 3D models simply by entering text prompts. It also provides robust image-to-3D conversion, which can transform 2D reference images into fully realized, three-dimensional assets. Beyond basic generation, TRELLIS introduces advanced editing features that distinguish it from most competing models. It can generate coherent variants of an existing 3D asset based on text prompts, enabling users to quickly explore different design options such as changing materials, colors, or overall styles. Additionally, it supports precise local manipulation of 3D assets, allowing users to remove specific parts, add new components, or replace entire sections of a model without altering the rest of the asset. Finally, TRELLIS features flexible multi-format output, decoding its unified latent representation into three major industry-standard 3D formats: 3D Gaussians, Radiance Fields, and polygon meshes, which can be directly downloaded as GLB files for immediate use in downstream applications.

What TRELLIS Can Help You Achieve

TRELLIS revolutionizes 3D asset creation by eliminating the need for labor-intensive manual modeling. It allows users to generate high-quality 3D assets with intricate geometry and vivid textures in minutes, drastically reducing production time and associated costs. Its multi-format output capability ensures that generated assets are compatible with a wide range of 3D software and platforms, eliminating the need for time-consuming format conversion. The advanced editing features enable rapid design iteration: users can easily adjust the appearance of an asset or modify specific parts to meet their exact requirements without starting the creation process from scratch. Furthermore, TRELLIS-generated assets can be seamlessly composited together to create complex and vibrant 3D art designs, making it a powerful tool for both inpidual creators and large professional teams.

Pricing and Availability

TRELLIS is currently a pure research project focused on advancing the field of 3D generation technology. The official interactive demo hosted on its website is available for free public access, allowing anyone to test its text-to-3D and image-to-3D capabilities without any cost or registration requirements. The project team has officially announced that the full source code, pre-trained models (including the 2-billion-parameter flagship version), and the associated large-scale 3D asset dataset will be fully released to the public. Upon release, all these resources will be available for free non-commercial and academic research use. Official guidelines and licensing terms for commercial use will be published alongside the full release of the project, providing clear frameworks for businesses and professional users.

How to Use TRELLIS

Using TRELLIS is intuitive and accessible through its official online demo. First, navigate to the official website at https://trellis3d.github.io. Locate and click on the Demo section to enter the interactive generation interface. Next, select your preferred generation mode: choose “Text to 3D Asset” to create a model from a written description, or “Image to 3D Asset” to convert a 2D image into a 3D model. For text-to-3D generation, enter a detailed and specific text prompt that describes the object you want to create, including its shape, material, color, texture, and any distinctive features. For image-to-3D generation, upload a clear, high-resolution reference image with minimal background clutter and no significant occlusions to ensure the best results. Once you have submitted your prompt or image, wait for the model to process and generate the 3D asset. After generation is complete, you can preview the 3D model directly in your browser and download it in the widely compatible GLB format for further editing or integration into your projects. For editing functions, upload an existing 3D asset, enter a text prompt specifying the desired changes (either variant generation or local manipulation), and the model will generate the edited version for you.

Who Should Use TRELLIS

TRELLIS is designed for a broad spectrum of users involved in 3D content creation, research, and design. It is an invaluable tool for 3D artists and digital designers who want to accelerate their workflow and explore new creative possibilities without being limited by manual modeling skills. Game developers can use it to quickly generate a large volume of perse game assets such as characters, weapons, vehicles, and environmental elements, significantly reducing development cycles. Animation studios and VR/AR content creators can leverage its high-quality output to produce immersive 3D content for films, animations, virtual showrooms, and interactive training simulations. Product designers can use TRELLIS to create rapid 3D prototypes of new products, facilitating early design reviews and iterations. Academic researchers in computer vision and computer graphics can use the open-source code and models to advance their own research in 3D generation and related fields. Additionally, it is accessible to digital art enthusiasts and hobbyists who want to create 3D content without extensive professional training.

Applicable Scenarios

TRELLIS finds practical applications across numerous industries and creative disciplines. In the gaming industry, it is ideal for generating a wide variety of game assets, from small props and characters to entire architectural environments. For animation and visual effects production, it can be used to create preliminary 3D assets for storyboarding and pre-visualization, as well as final assets for less critical scenes, significantly speeding up production pipelines. In the VR and AR space, TRELLIS enables the rapid creation of 3D content for virtual experiences, educational simulations, and digital twins. Product design teams can use it to generate 3D prototypes of consumer electronics, industrial equipment, and consumer goods, allowing for faster design validation and iteration. Digital artists can use TRELLIS to create unique 3D art pieces, digital sculptures, and NFT collections. It is also highly useful for educational purposes, helping educators create interactive 3D models for STEM courses and other subjects. Furthermore, the generated mesh assets can be directly used for 3D printing, turning digital designs into physical objects.

Advantages Over Competing Methods

TRELLIS offers several significant advantages over existing 3D generation models, positioning it as a leading solution in the field. First and foremost, its unified Structured Latent (SLAT) representation is a transformative innovation that allows a single model to output multiple high-quality 3D formats, including 3D Gaussians, Radiance Fields, and meshes. This eliminates the need to use separate models for different output types and simplifies the entire asset creation pipeline. Second, TRELLIS uses rectified flow transformers specifically adapted to handle the sparsity inherent in SLAT, and has trained models with up to 2 billion parameters on a large dataset of 500,000 perse 3D objects. This large-scale training results in significantly higher generation quality compared to most existing methods of similar scale, with more detailed geometry and more realistic textures. Third, TRELLIS introduces advanced 3D editing capabilities that were not available in previous mainstream models. Its ability to generate coherent asset variants and perform precise local manipulation allows users to easily refine and customize their 3D models, greatly enhancing creative flexibility. Fourth, it supports both text and image conditioning, providing users with multiple ways to input their creative ideas and adapt to different workflow preferences. Finally, the generated assets are of sufficient quality to be used directly in most downstream applications with minimal post-processing, saving users valuable time and effort.

Usage Tutorials and Reviews

Official step-by-step tutorials for all TRELLIS features, including text-to-3D generation, image-to-3D conversion, asset variant generation, and local manipulation, are available on the official website. These tutorials provide clear instructions and best practices to help users optimize their prompts and get the best possible results from the model. Independent third-party reviews and benchmarks have consistently praised TRELLIS for its exceptional generation quality, versatility, and editing capabilities. Evaluations have shown that TRELLIS outperforms many popular 3D generation models in terms of geometric accuracy, texture detail, and overall visual fidelity, particularly when generating complex objects such as mechanical devices, architectural structures, and detailed sculptures. Reviewers have also highlighted the convenience of its multi-format output and the significant efficiency gains provided by its editing features. However, reviews also note that TRELLIS currently has limitations in generating highly photorealistic real-world objects, and it performs best when creating artistic and stylized 3D assets.

Important Considerations When Using TRELLIS

There are several important points to keep in mind when using TRELLIS to ensure optimal results and responsible use. First, TRELLIS is currently optimized for generating artistic and stylized 3D assets, and its ability to produce highly photorealistic representations of real-world objects is limited. Users should adjust their expectations accordingly when generating realistic objects. Second, the quality of the generated results is highly dependent on the clarity and specificity of the input prompts or reference images. For text prompts, include as much detail as possible about the object’s shape, material, color, texture, and desired style. For image inputs, use high-resolution images with even lighting, no significant occlusions, and a plain background. Third, the GLB files available for download are created by baking the appearance information from 3D Gaussians onto polygon meshes. For very complex assets, minor artifacts may be present, and some additional post-processing in dedicated 3D editing software may be required. Fourth, as a research project, TRELLIS may occasionally produce incomplete or imperfect results. It is recommended to generate multiple versions of important assets and select the best one. Finally, users must adhere to responsible AI principles when using TRELLIS. The datasets used to train TRELLIS are publicly available and have been reviewed to exclude personally identifiable information and harmful content, but potential biases may still exist due to their internet-sourced nature. Do not generate 3D assets that contain harmful, offensive, or copyrighted content, or that violate the privacy of inpiduals.

Frequently Asked Questions

Is TRELLIS available for commercial use?

TRELLIS is currently a pure research project. Upon its full release, the code, pre-trained models, and datasets will be available for free non-commercial and academic research use. Official commercial use policies and licensing terms will be announced at a later date when the project is fully launched.

What output formats does TRELLIS support?

TRELLIS can decode its unified structured latents into three primary 3D representations: 3D Gaussians, Radiance Fields, and polygon meshes. Users can directly download generated assets in the widely compatible GLB format from the official demo.

Can I edit my existing 3D assets with TRELLIS?

Yes, TRELLIS provides two powerful editing features. Asset variant generation allows you to create different versions of an existing 3D asset by changing its material, color, style, or other attributes using text prompts. Local manipulation enables you to make targeted changes to specific regions of an asset, such as removing parts, adding new components, or replacing entire sections.

What are the main limitations of TRELLIS?

The main limitation of the current version of TRELLIS is its relatively limited capability in generating highly photorealistic real-world objects. It excels at creating artistic and stylized 3D assets. Additionally, extremely complex or highly detailed objects may require some post-processing to achieve perfect results.

Do I need a powerful computer to use TRELLIS?

For basic use through the official online demo, no special hardware is required, as all processing is done on cloud servers. If you plan to run TRELLIS locally after the code release, you will need a computer with a dedicated GPU with sufficient VRAM. A GPU with at least 16GB of VRAM is recommended to run the 2-billion-parameter model smoothly.