IndexMapAnti-DoomscrollAbout
Model FamilyCross-Modal

TRELLIS

Also known as: Microsoft TRELLIS · TRELLIS.2 · TRELLIS.2-4B · TRELLIS 2

Plain English

Microsoft Research's TRELLIS family generates 3D assets from prompts. TRELLIS.2 takes a single image and generates a textured mesh with materials that respond to scene lighting, with a .glb exporter. Each result estimates the object's hidden surfaces and can need mesh cleanup.

Technical Overview

TRELLIS.2 uses flow-matching transformers to generate a sparse voxel layout, then geometry latents, then material latents conditioned on the image and geometry. A sparse 3D variational autoencoder compresses O-Voxel data by 16× along each spatial axis. Cascaded inference can raise the output grid to 1536³, with more compute for fine detail.

Timeline (4)

RSS
  1. 2024

    Paper

    Xiang and colleagues introduced TRELLIS and its Structured Latent (SLAT) representation. A sparse 3D grid anchors features fused from multiple views, letting the same latent decode into a mesh, a radiance field, or 3D Gaussians. Rectified-flow transformers generate those latents from text or image conditions.

  2. 2025

    Release

    Microsoft Research released TRELLIS.2-4B weights and image-to-3D inference code under the MIT license. The model generates geometry and physically based rendering (PBR) attributes, including base color, roughness, metallic, and opacity. Microsoft reports shape and material generation times of about 3 seconds at 512³, 17 seconds at 1024³, and 60 seconds at 1536³ on an NVIDIA H100. The documented local setup uses Linux and an NVIDIA GPU with at least 24 GB of VRAM. Raw meshes can contain small holes. The supplied GLB exporter is a separate step and sets materials to opaque by default, so viewers render them as opaque even when the baked texture has an alpha channel.

  3. 2026

    Release

    Microsoft published TRELLIS.2's training code, with configurations for the sparse autoencoders and flow models. Training or fine-tuning on a custom 3D asset library starts by converting its meshes and materials into O-Voxel data.

See also (7)

Open full graph
2 more connections