We ablate our SDS-adaptive weighting mechanism by replacing all adaptive regularization weights with constant values across both physics guidance losses and appearance regularization (as shown in Fig. 8). We evaluate the following variants:
- λ=1 - very low constant weight produces severe visual artifacts and noise due to insufficient regularization, resulting in unstable optimization.
- λ=100 - low constant weight shows reduced but still significant noise and color distortions throughout the scene.
- λ=104 - moderate constant weight exhibits unnatural color artifacts with pinkish tones.
- λ=106 - extremely high constant weight over-regularizes the optimization, preventing refinement and producing results where minimal optimization occurs.
| λ=1 | λ=100 | λ=104 | λ=106 | Ours |
|---|---|---|---|---|
|
|
|
|
|
|