Microsoft Research just dropped TRELLIS.2, and it quietly moves the goalposts for image-to-3D. Feed it one picture and it hands back a 3D asset at up to 1536³ resolution with real PBR materials — base color, roughness, metallic, opacity — in seconds. The trick is a new way of storing 3D, and it might be the most important idea in AI 3D this year.
The Story
The first TRELLIS was a 2025 hit. It turned one image into a mesh, a Gaussian splat, or a radiance field in about 30 seconds, and it was MIT-licensed. Half the indie 3D pipelines you saw last year had TRELLIS somewhere inside them.
TRELLIS.2 rebuilds the core. The old model leaned on a sparse voxel grid plus separate decoders. The new one introduces the O-Voxel (Omni-Voxel): a single “field-free” structure that stores geometry and appearance together. Think of it as one grid that already knows both the shape and the surface, instead of stitching those together at the end.
Why that matters: the O-Voxel handles messy real geometry. Open surfaces, non-manifold edges, hidden internal parts — the stuff that made older generators blur into a melted potato. It uses “Flexible Dual Grids” to keep sharp edges crisp instead of rounding everything off.
The numbers are the headline. It is a 4-billion-parameter model. On an H100 it generates a 512³ asset in 3 seconds, 1024³ in 17 seconds, and a full 1536³ asset in about 60 seconds. A Sparse Compression VAE squeezes a 1024³ scene down to roughly 9,600 tokens, which is how they keep it fast. Converting a mesh back into an O-Voxel takes under 10 seconds on CPU, and going the other way runs in under 100 milliseconds on CUDA.
Why You Should Care
Most AI 3D tools still stop at “a shape with a baked-on color.” That looks fine in the demo and falls apart the moment you drop it into a scene with real lighting. TRELLIS.2 outputs proper PBR channels, so the asset reacts to your lights and reflections like a hand-authored model would.
The other quiet win is the two-way street. Because meshes convert into O-Voxels and back cleanly, this is not just a one-shot generator — it is a representation you could edit, re-texture, or feed into a bigger pipeline. That is the direction serious 3D tooling needs to go.
Now the honest caveat. The original TRELLIS was MIT and free for commercial use. TRELLIS.2 ships under a research-only license — academic and research use, not production shipping. The code, the demo, and a 500K-item dataset are public, so you can absolutely study it and prototype. Just do not build a paid product on it until the license changes. Keep an eye on the repo; these things sometimes loosen up.
Try It / Follow Them
- Project page: microsoft.github.io/TRELLIS.2
- Live demo: Hugging Face Space
- Code: github.com/microsoft/TRELLIS.2
- Paper: arXiv 2512.14692
The fastest test: open the Hugging Face Space, drop in a product photo or a character render, and watch what comes back. Then rotate it under different lighting to judge those PBR maps.
IK3D Lab Take
The O-Voxel is the real story here, not the token count. For two years the whole field argued about output formats — mesh vs. splat vs. NeRF — as if you had to pick a team. TRELLIS.2 says: store it once, in a structure that respects real geometry, and decode to whatever you need. If that idea sticks, it becomes the plumbing under a lot of tools you already use. Research license today, but this is the paper to read this month.
