Nikolaos Sarafianos

Our novel representation makes it effortless to leverage foundational 2D models, such as diffusion models, to directly generate 3DGS.

Work with colleagues at Brown and Meta Reality Labs

UVGS enables the application of image-based models for 3DGS feature extraction by representing it as multi-channel images.

We demonstrate effective compression of heterogeneous UVGS features and successful generalization of pre-trained 2D VAEs to this new representation.

𝗜𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗶𝗻𝗴 𝗨𝗩𝗚𝗦

We introduce 𝗨𝗩𝗚𝗦, a new 2D representation of 3D Gaussian Splatting (3DGS) that leverages spherical mapping.

Website: aashishrai3799.github.io/uvgs
Paper: arxiv.org/abs/2502.01846

Thank you @miniemeyer.bsky.social for sharing this, lots of familiar folks in there.

@chrisoffner3d.bsky.social if you could kindly add me to the list as well it’d be lovely. I spent most of my time on 3D generative models these days :)

At test-time given a 3D mesh, some image guidance with the desired edit, and a coarse 3D mask, it generates the edited 3D mesh in seconds.

Work with William Gao, Dilin Wang, Yuchen Fan, Aljaž Božič, Tuur Stuyck, Zhengqin Li, Zhao Dong and Rakesh Ranjan

By leveraging the representation power of large reconstruction models, we've developed a new 3D-consistent multi-view masking strategy that allows us to train our model to learn to fill in missing 3D information.

A New Approach for 3D Mesh Editing

Adding rabbit ears to a bird in 3D is a pretty challenging task but it's now possible with our new approach.

Website: chocolatebiscuit.github.io/MaskedLRM/
Arxiv: arxiv.org/abs/2412.08641