I'm surprised at the lukewarm reception. Admittedly I don't follow the image-to-3D space as much, but last time I checked in, the gloopy fuzzy outputs did not impress me.

I want to highlight what I believe is the coolest innovation: their novel O-Voxel data structure. I'm still trying to wrap my head around how they figured out the conversion from voxel-space to mesh-space. Those two worlds don't work well together.

A 2D analogy is that they figured out an efficient, bidirectional, one-shot method of converting PNG's into SVG's, without iteration. Crazy.

As an old guy

"System: The code is currently tested only on Linux."

:)

TRELLIS 1 had a massive impact on the research in this area, not least because it’s actually open (full dataset, training and inference). Research like SynCity or PhysX-3D (not the NVIDIA one) wouldn’t have been possible.

Excites for the follow ups for this new generation.

State of the art for Open Source. This is a nice improvement, but far out I cannot wait for a Sparc3D equivalent for local use. Its a step change in quality. I really hope Hunyuan3D-3 is the one to level up to that quality now

Where they get training data from?

You can play with a demo here: https://huggingface.co/spaces/microsoft/TRELLIS.2 (requires a hugging face acct)

The results from arbitrary pictures are not nearly as good as what's shown in the posting. So either the demo is running a gimped version of the model or the examples are _very_ handpicked.

Needs 24GB gpu to run.

Project website gives a nice look: https://microsoft.github.io/TRELLIS.2/