How TRELLIS.2 Works: O-Voxel, Structured Latents, and PBR Output

Dec 24, 2025

TRELLIS.2 converts one image into a 3D asset by combining image conditioning, sparse structure generation, compact native 3D latents, geometry generation, and material generation. Microsoft describes the released model as a 4B-parameter image-to-3D system capable of producing assets from 512³ to 1536³ resolution.

This article sticks to the published project and paper. It does not invent face counts, hidden-side accuracy, consumer-GPU timings, or a smaller unpublished model.

1. The image is a condition, not a complete 3D scan

A single image provides visible color, silhouette, lighting, and material cues. It does not directly reveal the back or every occluded surface. TRELLIS.2 uses the image to condition generation, but unseen geometry remains an inference.

That is why a clean, centered image with one object and limited occlusion is more useful than a crowded scene.

2. O-Voxel represents geometry and appearance

At the core is O-Voxel, a field-free sparse voxel structure. Microsoft states that it can represent:

  • open surfaces
  • non-manifold geometry
  • enclosed internal structures
  • Base Color
  • Roughness
  • Metallic
  • Opacity

This native geometry-and-appearance representation is designed to avoid lossy conversion through an iso-surface field.

Read the O-Voxel explanation.

3. A sparse 3D VAE creates compact structured latents

TRELLIS.2 uses a Sparse 3D VAE with 16× spatial downsampling. The goal is to preserve native 3D structure while making the latent representation compact enough for large generative models.

“16× spatial downsampling” is the project's published figure. It should not be rewritten as an unsupported token count or a made-up compression ratio for every asset.

4. Large flow-matching models generate structure, shape, and texture

The released training configuration and paper describe large flow-matching models for sparse structure, shape latents, and texture latents. In practical terms:

  1. infer where the asset's sparse 3D structure should exist
  2. generate shape information within that structure
  3. generate appearance/PBR attributes conditioned on the image and shape

This staged process separates coarse spatial organization from detailed geometry and appearance.

5. Post-processing creates a usable mesh and GLB

The official code demonstrates:

  • mesh simplification
  • optional remeshing
  • decimation target
  • texture-size selection
  • UV/material post-processing
  • GLB export

The web generator exposes a subset of those controls and provides browser preview and download.

Official performance figures

Microsoft reports these H100 measurements:

ResolutionTotalShape + material
512³about 3s2s + 1s
1024³about 17s10s + 7s
1536³about 60s35s + 25s

These figures exclude a hosted service's queue, upload, download, and browser time. They should not be generalized to other GPUs.

What to inspect in the result

Before calling an asset production-ready:

  • rotate behind it and inspect inferred surfaces
  • use wireframe to inspect density and problematic edges
  • verify UVs and PBR maps
  • check scale and axes after import
  • test transparency explicitly
  • measure cleanup time

Primary sources

Try the workflow

Generate a TRELLIS 2 GLB, then use the Blender and engine export guide to validate the file.

trellis2.com

trellis2.com

How TRELLIS.2 Works: O-Voxel, Structured Latents, and PBR Output | trellis2.com