Video
Inspect grounded tokens, Gaussian groups, and attention in 3D.
Explore exported scenes with interactive controls for query selection, visualization mode, and local structure.
Anchor-centered Gaussian group
Current selection
Query anchor 666 is highlighted in red. Child Gaussians appear in warm orange, while the context cloud stays in low-opacity gray.
Interaction hint
Use mode switches to change the emphasis. "Anchor → GS" is the default because it best communicates LocusGS design.
TokenGS vs LocusGS
Compare token-level Gaussian organization directly in 3D. By default, both panels show only the full Gaussian cloud. Enter a query index to highlight the corresponding Gaussian group.
TokenGS
All TokenGS Gaussians stay visible as the global organization baseline. Query-specific GS are highlighted in orange.
LocusGS
All LocusGS Gaussians stay visible as the global organization baseline. Query-specific GS are highlighted in orange.
Rendering & Depth Comparison
Switch scenes below and compare the two methods across RGB rendering and depth in a fixed 2 by 2 matrix.
Training Convergence
We further visualize the optimization dynamics on both datasets. Each block groups the two core training-time metrics and shows how LocusGS and TokenGS evolve over epochs.