Spaces:
Running
Trellis VS Trellis 3DGS
So i notice these two different models in the arena, and the 3DGS variant seems to output superior results, at least in my opinion, compared to regular Trellis. It's also currently winning the leaderboard. However, i couldn't find any information about "3DGS" online, and both these models just link directly to the same Trellis huggingface page, and there wasn't any info there either.
Can someone please explain to me the difference between these two models, and perhaps provide me with 'any' info about what 3DGS is and where to get it etc. Would be appreciated. TIA.
3DGS stands for Gaussian Splatting and is a way to represent 3D shapes (https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/, https://huggingface.co/learn/ml-for-3d-course/en/unit3/introduction). TRELLIS comes with 3 pretrained decoders to generate meshes, 3D Gaussian Splatting and Radiance Fields from the structured latents their flow transformer generated. This meams the "models" are the same until the last decoder (see also the image from the paper). You can also check their example implementation. They explain it in the code (https://github.com/microsoft/TRELLIS/blob/main/example.py):
outputs is a dictionary containing generated 3D assets in different formats:
- outputs['gaussian']: a list of 3D Gaussians
- outputs['radiance_field']: a list of radiance fields
- outputs['mesh']: a list of meshes
If you just want to generate the 3DGS you can call pipeline.run(image, formats=['gaussian']). Hope this helps.