1. Differentiable Volume Rendering
256 × 256 images, 64 samples per ray. The depth map is normalized by its maximum value.





2. Optimizing a Basic Implicit Volume
1000 epochs, Adam learning rate 0.0005, RGB MSE. Final parameters, rounded to two decimal places:
| Parameter | x | y | z |
|---|---|---|---|
| Center | 0.25 | 0.25 | 0.00 |
| Side lengths | 2.01 | 1.50 | 1.50 |

The recovered box has the expected longer x dimension and displaced center, consistent with the reference rendering.
3. Neural Radiance Field
Lego, 128 × 128, 250 epochs, 1024 rays per step, 128 samples per ray, initial learning rate 0.0005. The position-only MLP uses positional encoding and RGB MSE.

4.1 View Dependence
Materials, 128 × 128, 250 epochs, initial learning rate 0.0001. Both runs use 1024 rays per step and 128 samples per ray. Viewing direction is added only to the color branch; density depends on position.


Direction input changes brightness and highlight contrast on some spheres. It can represent reflections that a single color per point cannot, but extra directional capacity can overfit training views. Keeping density independent of direction and using a small color branch limits this risk. The visual difference is modest at this resolution.
4.3 VolSDF
Lego, 128 × 128, 250 epochs, initial learning rate 0.0005, 128 samples per ray, α = 10, Eikonal weight = 0.02. I compare three β values with the other settings fixed.
| β | Observation |
|---|---|
| 0.01 | Large floating structures and background clutter. |
| 0.05 — selected | Best visual balance of color and shape; background fragments remain. |
| 0.10 | Softer color and incomplete geometry. |


α controls interior density and opacity. β controls the width of the density transition around the zero-level surface.
- High vs. low β: High β favors a broad, diffuse boundary; low β concentrates density changes near the surface and favors a sharper boundary.
- Which is easier to train? High β gives gradients over a wider region. Low β has a narrow transition that fixed ray samples can miss.
- Which gives a more accurate surface? Low β can localize the surface better when sampling and optimization are adequate. Here, 0.01 produced more artifacts, so I select 0.05. The selected mesh still contains unwanted geometry.
5. 3D Gaussian Rasterization
The four provided unit tests pass. Pretrained chair, 256 × 256, 2048 Gaussians per splat. Each frame shows RGB, depth, and silhouette from left to right.

6. Training 3D Gaussian Representations
Truck, isotropic Gaussians initialized from the supplied point cloud, 1000 iterations, Adam, L1 RGB loss.
| Mean LR | Log-scale LR | Opacity-logit LR | Color LR |
|---|---|---|---|
| 0.0001 | 0.005 | 0.05 | 0.01 |
| Test PSNR | Test SSIM |
|---|---|
| 29.438 dB | 0.938 |


7.1 Rendering Using Spherical Harmonics
I render the pretrained degree-3 SH coefficients. The DC-only GIF is shown in Part 5.

Same camera in each pair. Left: DC only. Right: full SH.


