TRELLIS 2 is here: Microsoft's TRELLIS.2-4B, explained

Microsoft Research has released TRELLIS.2-4B, the successor to TRELLIS — the open-source model that powers most affordable image-to-3D services, including Pic2Mesh. Like the original, it turns a single image into a textured 3D asset, and like the original it ships under the MIT license. Under the hood, though, it is a substantial redesign rather than a fine-tune.

Everything below comes from the official Hugging Face model card — no speculation. And to be upfront: Pic2Mesh runs TRELLIS 1 today. More on that below.

What's new in TRELLIS 2 vs TRELLIS 1

The original TRELLIS generates in a structured latent (SLAT) space — a sparse 3D grid of latent features decoded into a mesh, radiance field, or Gaussians, with textures baked from the input photo. TRELLIS 2 replaces that pipeline with what Microsoft calls O-Voxel: a “field-free” sparse voxel structure that encodes geometry and appearance together, generated by a 4-billion-parameter flow-matching transformer. Per the model card, that buys:

• Arbitrary topology — open surfaces and non-manifold geometry no longer break the reconstruction, a known weak spot of iso-surface approaches like SDF or Flexicubes
• Full PBR materials — not just baked color textures but physically-based material attributes, including opacity for transparent and translucent surfaces
• Higher resolution — generation at voxel grids from 512³ up to 1536³, with a sparse 3D VAE that compresses a 1024³ asset into roughly 9.6K latent tokens
• Speed — the card reports about 3 seconds per object at 512³ on an NVIDIA H100, ~17 seconds at 1024³, and ~60 seconds at 1536³
• Shape-conditioned texturing — it can generate textures for an existing mesh from a reference image, not only full generations from scratch

The card is also candid about limitations: raw meshes can contain small holes (Microsoft ships hole-filling post-processing scripts for 3D-printing use), and TRELLIS.2-4B is a base model with no preference alignment, so output style tracks the training data.

Does Pic2Mesh run TRELLIS 2?

Not yet — honest answer. The generator on this site runs the original TRELLIS through our inference provider. That provider has begun listing a TRELLIS 2 endpoint, but at roughly 10× the per-generation cost of TRELLIS 1 — so we are testing output quality against that price before wiring it in, most likely for credit-pack users first. No fake “now with TRELLIS 2!” badge here until it is actually behind the button. This page will be updated when it is.

Turn an image into a 3D model right now

While TRELLIS 2 support settles, the TRELLIS 1 pipeline on Pic2Mesh already produces clean textured meshes from a single photo — and it is free to try:

• Open the generator and upload one JPEG, PNG, or WebP of a single object
• Wait about a minute while the model builds geometry and textures
• Inspect the result in your browser, then download a GLB that drops into Blender, Unity, Unreal, or your slicer

You get 3 free generations a day with no signup and no watermark — enough to judge the quality on your own photos.

Want to run TRELLIS.2 yourself?

The weights are open, so you can. Budget for it: the model card lists an NVIDIA GPU with at least 24 GB of memory (verified on A100 and H100), Linux as the only tested OS, and a CUDA 12.4 toolchain to compile dependencies. If that sounds like an afternoon you would rather not spend, browser tools like this one exist for exactly that reason.

FAQ