TRELLIS is an open-source image-to-3D generative model from Microsoft Research, with co-authors from Tsinghua University and USTC: it takes an image and generates a 3D asset, a textured mesh that 3D software can load. The version on the current project page, TRELLIS.2, has 4 billion parameters. Microsoft describes it as a research project.
TRELLIS.2 stores shapes in O-Voxel, a sparse voxel structure (a voxel is a small cube in a 3D grid; sparse means only occupied cubes are stored) that encodes both geometry and surface materials. A sparse compression VAE (variational autoencoder, a network that compresses data into a small latent code) with 16× spatial downsampling makes each asset compact, and diffusion transformers (DiTs) generate new assets in that compact space. The output can contain open surfaces and internal structures, and PBR materials (physically based rendering: base color, roughness, metallic and opacity), at resolutions up to 1536³. On an NVIDIA H100 GPU, a 1024³ asset takes about 17 seconds.
The code is on GitHub under the MIT license and is free to use; it is tested on Linux and needs an NVIDIA GPU with at least 24 GB of memory.