Users praised the TokenGS paper for its method of converting generated videos into explorable 3D Gaussian Splatting scenes, calling the work great.
We introduce three new techniques: latent bottlenecks, scene-latent tuning, and mean-of-gradients training.
Together, they improve both reconstruction quality and training efficiency.
More details in the release note: https://github.com/nv-tlabs/TokenGS/blob/main/release_2026.6.md
All models and code are now open-sourced!
Check out the GitHub repo: https://github.com/nv-tlabs/TokenGS (4/4)
We’ve also released the Kubric model, showcasing TokenGS’s dynamic reconstruction capability and static-dynamic decomposition.
@jiawei6_ren It's a great paper, nice work