NTF has the highest PSNR by 0.29 dB, while SVF has slightly better SSIM and LPIPS. Both substantially outperform Vanilla 3DGS in this mixed-architecture comparison.
The fixed 10 dB NTF gate removes useful gradients even at moderate noise. SVF retains sharper facades, cars and tree structure and gains 3.36 dB over NTF.
NTF begins to smear the roofline and expand the distant tree into a blob. SVF wins all three metrics and improves 2.36 dB over Vanilla.
With the corrected manifest calibration, SVF beats NTF by 0.93 dB and Vanilla by 2.83 dB. The tree is still imperfect, confirming that loss design alone does not eliminate the representation bottleneck.
The fixed gate is not yet catastrophic at this SNR. NTF wins all three unaligned metrics, while SVF stays between NTF and Vanilla.
NTF has a 0.42 dB PSNR edge; SVF has slightly better LPIPS. Both improve upon Vanilla, suggesting noise-aware supervision remains useful.
SVF improves 1.37 dB over NTF and 1.10 dB over Vanilla, with matching gains in SSIM and LPIPS. This is the strongest real-data support for threshold-free stabilization.
SVF is better than NTF in PSNR and SSIM, but worse in LPIPS—and Vanilla 3DGS is stronger overall. Keeping every stabilized residual is insufficient when observations are almost entirely unreliable.
