TRELLIS

TRELLIS, from Microsoft Research, generates 3D assets from text or image prompts and can decode a single generation into radiance fields, 3D Gaussians, or textured meshes. That flexibility comes from its Structured LATent representation, SLAT, a unified intermediate format that captures both geometry and appearance and converts between output types without regenerating. Rectified flow transformers operate on this representation, with pretrained models ranging from 342 million to 2 billion parameters trained on a curated set of 500,000 diverse 3D objects. Released variants include TRELLIS-image-large at 1.2 billion parameters plus text-conditioned models in base, large, and extra-large sizes, and the system supports local editing to produce variants or modify specific regions of an asset. The code is MIT licensed, with a few rendering components retaining their upstream licenses. Game developers, technical artists, and researchers use it for asset prototyping, product visualization, and image-to-3D research pipelines.