Trellis (service for creating 3D models from photos)

History

2024: Product Announcement

In early December 2024, Microsoft introduced a new artificial intelligence-based service, the Trellis platform. This open source neural network is designed to generate 3D models of objects from a photo or text description.

It is noted that Trellis can form 3D models of complex objects with many details. In this case, the created options can be edited and changed: in particular, you can ask the neural network to add any elements through text queries, modify the texture, etc. According to the developers, the new service is superior in capabilities to other similar solutions.

𝓲

Microsoft

The neural network is based on the Structured LATent (SLAT) method, which decodes input data into various formats, including polygonal grids. The training used information on 500 thousand various objects. The Trellis service is available on the Hugging Face platform.

Introducing a new method for creating 3D structures. Our model generates high-quality results based on text or graphic conditions, significantly superior to existing methods, the project website says.

The Trellis AI platform can be used in two modes - via the Internet and locally. In the first case, it is possible to generate 3D objects only by image - you can use your own originals or samples prepared by the developers. With a local deployment, text queries become available, but a sufficiently powerful graphics accelerator with 16 GB of video memory is required. When working with images, the system removes the background and then generates a 3D model, which can be adjusted with tooltips if necessary. The finished result is available in GLB format for use or installation in Blender or another application.^[1]

Notes