TRELLIS.2 4B: Model Specs, VRAM, and Output Limits

What the 4B model is

TRELLIS.2 is Microsoft's 4-billion-parameter image-to-3D model. Its published design uses O-Voxel, a sparse voxel representation intended to handle open surfaces, non-manifold geometry, and internal structure without a lossy conversion step. The result is a textured mesh, not a CAD solid.

The distinction matters when you choose a workflow. A 4B model can produce a useful starting asset from one image, but it cannot recover information the image never contained. The back, underside, internal cavities, and true scale are estimates until you inspect and correct them.

Published output range

The official project documents:

• Image-to-3D generation at 512³, 1024³, and 1536³
• PBR materials, including base colour, roughness, metallic, and opacity
• GLB export through the project's post-processing
• Parameters for decimation and texture size

The hosted service exposes the same practical shape: one source image goes in, one textured GLB draft comes out. It does not add a rig, animation, or engineering tolerance to the result.

Hardware and timing

The official repository currently calls for Linux, an NVIDIA GPU with at least 24 GB of memory, and verification on A100 and H100 cards. The documented environment uses Python 3.8 or later, PyTorch 2.6.0, CUDA 12.4, and flash-attn with an xformers fallback.

Microsoft's published inference timings are measured on an H100. They are model runtime figures, not a promise about a consumer GPU, a laptop, or the complete time from upload to a reviewed asset.

What the GLB does not contain

• A skeleton, weight painting, or animation-ready topology
• Production UVs and a platform-specific polygon budget
• Real-world scale or mechanical tolerances
• A guarantee that thin walls are printable
• A licence for the source image or the depicted character

Use the GLB viewer to inspect the mesh, then use the 3D converter if your downstream tool needs a different format. For print work, follow the 3D printing workflow.

Hosted or local?

Use the hosted generator when you need to evaluate one image and do not want to manage CUDA, model files, and GPU memory. The selected resolution and credit cost are shown before submission.

Use a local install when you need data control, sustained volume, or internal integration. Read the download guide and the local setup checklist, and test the official pipeline before adding a community wrapper.

FAQ

Is TRELLIS.2 4B the same as the older TRELLIS?

No. The older TRELLIS release has different model names, setup commands, and output modes. Use the current repository and do not paste old commands into a TRELLIS.2 environment.

Does 4B mean the output is production-ready?

No. Parameter count describes the model, not the quality of a particular mesh. Topology, scale, materials, and the source image still need review.

Can I run 4B on a smaller GPU?

The official repository currently documents 24 GB or more of NVIDIA memory. A wrapper's low-memory option is not a guarantee; measure your own workload.

How do I inspect the result?

Download the GLB and open it in the free browser viewer. Orbit the model and check the sides the source image did not show.

Run a hosted 512 evaluation

Compare the official model with other open-weight releases