Training and rendering were performed on an NVIDIA H100 GPU through PSC Bridges-2.
| Parameter | Value |
|---|---|
| Box Center | (0.25, 0.25, 0.00) |
| Side Lengths | (2.01, 1.50, 1.50) |
| Setting | Value |
|---|---|
| Dataset | Lego |
| Image Resolution | 128 × 128 |
| Training Epochs | 250 |
| Rays per Batch | 1,024 |
| Samples per Ray | 128 |
| Position Harmonic Frequencies | 6 |
| MLP Layer Setting | 6 |
| Hidden Width | 128 |
| Input Skip Setting | Layer Index 3 |
| Initial Learning Rate | 0.0005 |
| Loss | Mean Squared Error |
| Setting | Value |
|---|---|
| Dataset | Materials |
| Image Resolution | 128 × 128 |
| Training Epochs | 250 |
| Rays per Batch | 1,024 |
| Samples per Ray | 128 |
| Position Harmonic Frequencies | 6 |
| Direction Harmonic Frequencies | 2 |
| Spatial Hidden Width | 128 |
| Direction Branch Hidden Width | 64 |
| Initial Learning Rate | 0.0001 |
| Loss | Mean Squared Error |
View dependence allows the network to represent reflections and other appearance changes across viewing directions. However, a highly expressive color branch can also memorize training views instead of learning an appearance model that generalizes.
I kept density independent of viewing direction and used fewer harmonic frequencies for direction than position. The direction branch also has a smaller hidden width. These choices limit its capacity, but a controlled evaluation on held out views would be needed to measure their effect on generalization.
The coarse network evaluates uniformly sampled points along each ray. Its rendering weights define a probability distribution for sampling additional points. These weights are detached before inverse CDF sampling.
The additional points are combined with the original samples and sorted by depth. The fine network evaluates this combined set. The training loss is the sum of the coarse and fine RGB reconstruction losses.
| Setting | Value |
|---|---|
| Dataset | Lego |
| Image Resolution | 128 × 128 |
| Training Epochs | 250 |
| Coarse Samples per Ray | 64 |
| Additional Fine Samples per Ray | 64 |
| Total Samples Evaluated by Fine Network | 128 |
| Rendering Chunk Size | 1,024 |
Hierarchical sampling directs additional samples toward regions that contribute strongly to the image. This can improve the sampling of surfaces without increasing the uniform sampling density everywhere.
However, it requires a second network and an additional rendering pass. The observed training rate was approximately 96 iterations per second, compared with approximately 175 iterations per second for the earlier Part 3 run. The architectures also differed, so this comparison does not isolate the cost of hierarchical sampling alone.
The resulting reconstruction remained consistent across views, but the visual improvement over the earlier NeRF was modest. The combined coarse and fine loss is not directly comparable to the single network RGB loss.
Alpha controls the density scale. Increasing alpha makes the interior more opaque over a given distance.
Beta controls the width of the density transition around the surface. A larger beta produces a broader transition, while a smaller beta produces a sharper transition near the zero level set.
High beta spreads density across a wider region around the surface. This can allow the model to explain images using a soft transition rather than a sharply localized boundary.
Low beta concentrates density changes near the surface. This supports sharper boundaries, but makes optimization more sensitive to the surface position and the sampled points.
A relatively high beta is generally easier during early training because a wider region provides useful rendering gradients.
With low beta, the transition becomes narrow and samples may miss it. This can make it harder to update an incorrectly positioned surface.
A suitably low beta can produce a more sharply localized surface once the geometry is reasonably initialized.
However, sufficient sampling and successful optimization are still necessary. A very low beta can hurt training if the transition is poorly sampled.
| Setting | Baseline | Selected Configuration |
|---|---|---|
| Alpha | 10.0 | 10.0 |
| Beta | 0.05 | 0.025 |
| Samples per Ray | 128 | 256 |
| Eikonal Weight | 0.02 | 0.02 |
| Training Epochs | 250 | 250 |
| Initial Learning Rate | 0.0005 | 0.0005 |
| Image Resolution | 128 × 128 | 128 × 128 |
The selected configuration produced cleaner color outlines and fewer floating streaks. Its geometry still contains rough thin structures and some merged components.
Reducing beta sharpens the density transition, while increasing the number of samples helps resolve that narrower transition. Since both settings changed together, the improvement should be attributed to the combined configuration.
I evaluated the Gaussian exponent using the displacement between each pixel and Gaussian center, together with the inverse covariance matrix.
The supplied unit tests reported:
4/4 Tests Passed
| Parameter | Learning Rate |
|---|---|
| Means | 0.001 |
| Scales | 0.005 |
| Colors | 0.01 |
| Opacities | 0.05 |
The smaller means learning rate limits large changes to the point cloud initialization. Colors and opacities can adapt more quickly to image supervision.
| Setting | Value |
|---|---|
| Optimizer | Adam |
| Training Iterations | 1,000 |
| Loss | L1 |
| Held Out Test Views | 5 |
Mean PSNR = 29.813 dB
Mean SSIM = 0.942
The top row shows predictions during training. The bottom row shows ground truth images.
With SH, the green backrest is darker and the gold floral pattern has stronger contrast. Without directional SH, the upholstery appears brighter and flatter.
With SH, the darker green regions contrast more strongly with the gold pattern. Streaks around the chair’s edges remain visible in both versions.
With SH, the central backrest becomes darker while the upper region remains brighter. This creates more variation in shading, while the chair’s outline remains unchanged.
| Setting | Baseline | Improved |
|---|---|---|
| Gaussian Count | 10,000 | 10,000 |
| Gaussian Type | Isotropic | Isotropic |
| Initialization | Random | Random |
| Initial Means Standard Deviation | 0.2 | 0.8 |
| Initial Scales | Starter random scales, approximately 0 to 0.01 | 0.03 |
| Initial Opacity | Approximately 0.9997 | 0.5 |
| Means Learning Rate | 0.0001 | 0.001 |
| Scales Learning Rate | 0.005 | 0.005 |
| Colors Learning Rate | 0.01 | 0.01 |
| Opacities Learning Rate | 0.05 | 0.05 |
| Training Iterations | 1,000 | 5,000 |
| Image Resolution | 128 × 128 | 128 × 128 |
| Loss | L1 | L1 |
| Optimizer | Adam | Adam |
The provided script evaluates these models on the held out validation split.
| Configuration | Mean PSNR ↑ | Mean SSIM ↑ |
|---|---|---|
| Baseline | 17.671 dB | 0.654 |
| Improved | 22.628 dB | 0.823 |
| Improvement | 4.957 dB | 0.169 |
I made the following changes compared with the baseline:
The broader initialization distributes Gaussians across a larger region. Larger initial scales increase their projected coverage, while lower initial opacity reduces the dominance of individual Gaussians during early training.
The higher means learning rate allows centers to move farther from their initial positions. Longer training provides additional opportunities to fit the scene.
The baseline contains blurred objects, incomplete coverage, and sparse detail. The improved configuration separates objects more clearly and reconstructs their colors and arrangement more successfully. Fine reflective details remain rough.
The improved configuration increased both PSNR and SSIM. However, several settings and the training duration changed together, and the runs did not use matched random initialization. Therefore, this comparison measures the combined configuration rather than the individual effect of each modification.
Both runs retained isotropic Gaussians and L1 loss. I did not use adaptive densification, anisotropic covariance optimization, or spherical harmonic coefficient training for this experiment.