Highest and lowest describe the model-guided result's final improvement over the baseline within the same run, using the reported PSNR, SSIM, LPIPS, and loss changes.

Interactive results

Comparison

Baseline vs model-selected

Metrics and Settings

Final checkpoint summary

The left side compares final baseline and model-selected metrics. The right side shows the selected multipliers applied to 3DGS after model testing.

Checkpoint Metrics

Training timeline

Checkpoint metrics compare baseline and model-selected reconstruction quality across training. Use the chart or data tab to inspect PSNR, SSIM, LPIPS, and loss over time.

Visual Comparison

Model previews

The left image shows the original reference crop. The right image lets you swipe between baseline and model-selected preview renders from the available final checkpoint.

Checkpoint Final checkpoint

Original reference preview — Original reference

Model-selected preview for swipe comparison

Baseline reconstruction Model-selected reconstruction

Data source View 3D

The swipe maps and preview renders show checkpoint outputs used for this research comparison. The 3D Gaussian Splatting models were stopped at 5,000 or 7,000 training iterations for controlled evaluation, so longer training would generally be expected to improve visual completeness and rendering quality.