TRELLIS.2-4B Guide 2026: Microsoft Image-to-3D Free Demo

Last updated: September 10, 2026

TRELLIS.2-4B is Microsoft's open-source image-to-3D model: feed it a single image and it produces a textured 3D mesh with full PBR materials, at up to 1536³ voxel resolution. It is one of the strongest open weights available for image-to-3D generation — and this guide covers everything you need to actually use it: what it is, what hardware it needs, how to run it in ComfyUI, and how to try it online without any GPU.

If you don't want to install anything, you can generate 3D models online directly in your browser — no Hugging Face account, no queue, no GPU required.

TRELLIS.2-4B at a Glance

Specs sourced from the official Microsoft project page and the Hugging Face model card.

What Is TRELLIS.2-4B?

According to Microsoft Research's official project page, TRELLIS.2-4B is an "open-source 4B-parameter image-to-3D model producing up to 1536³ PBR textured assets, built on native 3D VAEs with 16× spatial compression." It is the second generation of the TRELLIS ("Structured 3D Latents") family — the subtitle Native and Compact Structured Latents for 3D Generation describes the two core ideas: a native 3D representation not limited by iso-surface fields, and a compact latent space that makes large-scale generation practical.

Three technical components matter in practice:

1. O-Voxel (Omni-Voxel representation). Instead of predicting a continuous signed-distance field and extracting a mesh from it (the approach used by most image-to-3D models, which struggles with thin walls and open surfaces), TRELLIS.2 works directly on a sparse voxel grid. Geometry uses a Flexible Dual Grids scheme that preserves sharp edges, and the same voxel encodes appearance — so a single structure carries both shape and material.

2. SC-VAE (Sparse Compression VAE). A 3D VAE compresses the voxel data 16× per axis. A fully textured 1024³ asset shrinks to roughly 9.6K latent tokens with negligible quality loss — this is what makes a 4B transformer able to output high-resolution 3D at all.

3. Flow-matching transformer. The generative backbone predicts the structured latent in a flow-matching setup, which supports multi-resolution output and shape-conditioned texture generation (re-texturing an existing mesh while keeping its geometry).

TRELLIS.2 vs TRELLIS 1

If you used the original TRELLIS, the second generation is a significant upgrade rather than an incremental one:

For a deeper dive into the pipeline, see our guide on how TRELLIS 2 works.

What It Generates

High resolution with clean geometry. Sharp edges, thin geometry, transparent objects, open surfaces, and even interior structure survive generation — the classic failure cases of field-based models.

Production-oriented PBR. Output includes Base Color, Metallic, Roughness, and Alpha channels, so assets can drop into a PBR render pipeline without re-baking.

Flexible outputs. Multi-resolution generation lets you pick a trade-off between detail and file size, and shape-conditioned mode lets you keep a mesh you like and regenerate only its texture.

Three Ways to Use TRELLIS.2-4B

Option 1: Official Hugging Face demo (free, but limited)

Microsoft hosts a demo Space on Hugging Face. It's free, but requires a Hugging Face account and queues during peak hours, with limited generation time per session. Fine for a first test, impractical for production work.

Option 2: Run it locally (free, needs a GPU)

The weights are open, so you can run the full pipeline on your own hardware. The practical floor is around 6–8GB VRAM with optimized settings — see our TRELLIS 2 low VRAM guide for GPU benchmarks and the full installation walkthrough. Model files are on the TRELLIS 2 download page.

Option 3: Use it online, no GPU (fastest)

Trellis2.app runs image-to-3D generation on cloud GPUs: upload an image, get a PBR-textured GLB in minutes, with multi-view input support and no local setup. This is the same TRELLIS.2 pipeline without the hardware bill.

VRAM Requirements

Approximate requirements for local deployment:

The 4B transformer is the main memory consumer; the SC-VAE's compact latents keep the rest of the pipeline manageable. Detailed per-GPU benchmarks are in the low VRAM guide.

Using TRELLIS.2-4B in ComfyUI

ComfyUI integration is maintained via the Comfy-Org/TRELLIS.2 repackaged weights, so the model drops into standard node workflows. Our complete TRELLIS 2 ComfyUI guide covers custom node installation, workflow setup, and GLB export, and the workflow collection has ready-to-use graphs. If your exports are missing metallic or roughness channels, the ComfyUI PBR guide addresses exactly that.

From GLB to Blender, Unity, and Unreal

The fastest way to avoid "why does this look wrong in my 3D editor" is to validate the export immediately:

1. Check scale — generated meshes are not guaranteed to be real-world sized; normalize before placing in a scene.
2. Check texture channels — confirm Base Color, Metallic, Roughness, and Alpha maps are all wired in the material; alpha requires enabling transparency on the material.
3. Check orientation — up-axis mismatches (Y-up vs Z-up) are the most common import surprise.
4. Check topology density — 1536³ assets are dense; decimate before use in real-time engines if you need 60 fps.

License and Commercial Use

The TRELLIS.2 code and weights are publicly released by Microsoft Research. Note that the official project page describes the release as a research project, and the model license on Hugging Face governs actual usage terms — read the model card before shipping anything commercial. If licensing uncertainty is a concern, using a hosted service like Trellis2.app moves that responsibility to the platform.

FAQ

Is TRELLIS.2-4B free?

The weights and code are open source. Running it costs either your own GPU time or a hosted service's credits.

Where is the TRELLIS.2-4B download?

On Hugging Face (model weights) and GitHub (code). Our download guide walks through both.

Does it work with TRELLIS 2 ComfyUI workflows?

Yes — via the Comfy-Org repackaged weights. Start with the ComfyUI installation guide.

Is "Trellis 2.0" or "TRELLIS 3" a different model?

"Trellis 2.0" usually refers to the same TRELLIS.2 release. There is no TRELLIS 3 at the time of writing.

How is this different from Meshy or Tripo?

Those are closed commercial models. TRELLIS.2-4B is open-source, generally stronger on topology fidelity and PBR output, but needs more setup. For head-to-head details, see Meshy AI vs Trellis vs Tripo and our image-to-3D models on Hugging Face comparison.

Can I generate 3D without a GPU?

Yes — see how to use TRELLIS 2 online for free, or generate online and skip the hardware entirely.

Get Started

TRELLIS.2-4B is the current open-source reference for single-image 3D generation: native topology handling, real PBR output, and resolutions that finally hold up in a render pipeline. Try it in the free online 3D generator — no GPU, no account, no queue.