1.3

Left: Grid Visualization, Right: Ray Visualization

1.4

1.5

2.2

Box center: (0.2500840723514557, 0.2505473494529724, -0.00036280264612287283)

Box side lengths: (2.004598379135132, 1.5033210515975952, 1.5031969547271729)

2.3

3

4.1

After adding view dependence, the model gets slightly better performance than Q3's output. However, the difference is not as significant. In general, adding view dependence leads to better generalization and more realistic outputs than without it. This is because without having viewing directions as input, it won't be able to generate fine-grained specular effects.

4.3

Here is the output after training with the same network as Q3

Here is the output trained with double the number of sampling rays. Notice, that the quality increased only slightly

Here is the output trained with a deeper architecture: it uses 2 more hidden layers (8) and the dimension of MLP is also double (256). Notice, that this increased the quality considerably: notably the floor now has lot more detail than before.

Late Days: 0