Let it Snow! Animating 3D Gaussian Scenes with Dynamic Weather Effects via Physics-Guided Score Distillation

Supplementary Material

Please click on the sections below to view our interactive results and ablations:

`
• Our Results on Dynamic Weather Effects
• Our Results on Creative Text Prompts
• Comparison to 4D Editing Baseline
• Ablations
• SDS-Adaptive Weight Ablations
• Regularization Loss Ablations
• Trajectory Drift Ablations

Trajectory Drift Ablations

We investigate the trajectory drift threshold behavior discussed in our Limitations (Sec. 4.4) by evaluating constant regularization weights for position loss L_xyz (as shown in Fig. 10), replacing our SDS-adaptive approach. We measure Avg. Drift (mean Euclidean distance between rendered and simulated positions, shown in parentheses). We evaluate the following variants:
- λ_xyz=1 - insufficient regularization allows excessive drift beyond the reliable guidance threshold, causing active snow Gaussians to deviate from simulation and fail to reach the ground.
- λ_xyz=100 - low regularization still results in high drift, with snow trajectories deviating significantly from the physics prior.
- λ_xyz=10⁴ - moderate regularization reduces drift but trajectories still deviate noticeably from simulation.
- λ_xyz=10⁶ - excessive regularization minimizes drift but over-constrains optimization, preventing meaningful refinement.