Structured 3D Latents for Scalable and Versatile 3D Generation

https://github.com/microsoft/TRELLIS

TRELLIS: Structured LATent (sparse voxels with vision-foundation-model features); up to 2B parameters; ~500K 3D objects; decodes to radiance fields, 3D Gaussians, meshes; repo lists MIT license, 16 GB GPU, CVPR'25 spotlight.

Text- and image-to-3D asset generation