wtf is trellis?

TRELLIS is a research project from Microsoft that generates 3D objects from text descriptions or images. You describe something in words or provide a photo, and the model produces a three-dimensional version of it that can be exported in several common 3D formats used in games, film, and design software. The core idea is that TRELLIS uses a single internal representation to produce multiple output types from the same generation. Rather than committing to one specific 3D file format, it can output as a mesh (the kind of 3D model used in games and 3D printing), as a radiance field (a way of representing 3D scenes that captures lighting), or as 3D Gaussians (another representation suited to rendering). This flexibility is useful for teams with different downstream tools. The project provides pre-trained models ranging from 342 million to 2 billion parameters. Larger models generally produce more detailed results. A live demo is available on Hugging Face, so you can try the tool without installing anything. Image-based generation is recommended over text-based generation when you want the most detailed output, since the text-conditioned models have less training data to work with. Running the code yourself requires a Linux machine with an NVIDIA GPU that has at least 16 GB of memory. Installation involves setting up a Python environment with several GPU-specific libraries, which the project handles through a setup script. The training code is also publicly available for researchers who want to build on the work. The full README is longer than what was shown.