TokenGS

Decoupling 3D Gaussian Prediction from Pixels with Learnable Tokens

DL3DV Results

Interactive viewer for 6-view reconstruction on DL3DV (448x256 resolution).

Reference Image

RE10K Results

Comparison between our method and GS-LRM on 2-view reconstruction on RE10K (256x256 resolution). Note the GS-LRM artifacts visible in bird's eye view.

Reference Image

Test time training

Comparison between three test time training methods.

Reference Image

Scene Extrapolation

Comparison between our method and GS-LRM on scene extrapolation. Both methods are finetuned with extrapolation view sampling.
Left: GS-LRM. Middle: Ours. Right: GT.

Dynamic Reconstruction

Comparison between BTimer and our method on dynamic reconstruction. Left: BTimer.
Right: Ours.

BTimer

Ours

Emergent Scene Flow

Trajectories of the dynamic Gaussians across time.