1. 3D Gaussian Splatting

1.1 3D Gaussian Rasterization (35 points)

1.1.1-1.1.2

All unit tests passed

1.1.3-1.1.5

1.2 Training 3D Gaussian Representations (15 points)

Learning rates used: Opacities: 0.01 Scales: 0.01 Colors: 0.01 Means: 0.001 Number of iterations: 1000 PSNR: 29.857 SSIM: 0.943

1.3 Extensions (Choose at least one! More than one is extra credit)

1.3.1 Rendering Using Spherical Harmonics (10 Points)

For the follow visualizations, the first one will be from Q1.1, and the second one is from Q1.3.

In the GIFs below, we can see that the color in the Q1.3 output changes much more smooth, and specific parts of the chair change color. On the other hand, Q1.1's output has shading that looks less natural because it has a constant color.

Frame 001 In frame 001, we can see the effect of using the spherical harmonics by examining the seat of the chair. The color throughout the chair in Q1.3 is much smoother and follows looks like there is natural shading. On the other hand, Q1.1's output has a distinct boundary, where the top half of the seat is completely bright, and the lower half is completely dark.

Frame 005 In frame 005, we can see the effect of using the spherical harmonics by examining the seat of the chair again. In Q1.1's output, the color of the top half of the seat is the exact same as frame 001. It is completely bright. Q1.3's output shows that the color between frame 001 and 005 changes. The seat of the chair changes to a slightly brighter color.

2. Diffusion-guided Optimization

2.1 SDS Loss + Image Optimization (20 points)

Text prompt: "a hamburger" was run for 1600 iterations.

Text prompt: "a standing corgi dog" was run for 1500 iterations.

Text prompt: "a traffic cone" was run for 1000 iterations.

Text prompt: "a basketball player" was run for 1800 iterations.

2.2 Texture Map Optimization for Mesh (15 points)

Text prompt: "a gold and purple cow" Text prompt: "a muscular cow"

2.3 NeRF Optimization (15 points)

Prompt: "a standing corgi dog" Prompt: "a cow" Prompt: "a basketball". The shape of the basketball is clear and it is orange, but the lines that are usually in a basketball aren't very clear.

2.4 Extensions (Choose at least one! More than one is extra credit)

2.4.1 View-dependent text embedding (10 points)

Compared to the results from 2.3, we can see that the results below are much more consistent across different viewpoints. In both the cow and corgi gifs from Q2.3, we can see multiple front faces from different viewpoints. By adding in the view-dependent text embeddings, we are able to see that the results are more consistent. For example, a view on the back side of the corgi shows the backside of the corgi's head, instead of another face. We can see this in the cat video as well.

View-dependent "a standing corgi":

View-dependent "a cat":