16-825 Assignment 4

Andrew ID: rajathc

1. 3D Gaussian Splatting

1.1 3D Gaussian Rasterization (35 points)

1.1.1-1.1.2

[4/4] Tests Passed

1.1.3-1.1.5

1.2 Training 3D Gaussian Representations (15 points)

Number of training iterations: 1200

Training Progress:

Final Render:

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

1.3.1 Rendering Using Spherical Harmonics (10 Points)

Original Render (Q1.1.5):

Spherical Harmonics Render - View-dependent (Q1.3.1):

View Independent (Q1.1.5)	View Dependent (Q1.3.1)

Explanation of differences:
In the original render, the yellow or golden section of the chair maintains a consistent intensity across all viewpoints. However, in the spherical harmonics render, the color appearance changes with the viewing angle — for example, when viewed from directly above, it appears more golden, while from an oblique angle, the color becomes noticeably less intense.

2. Diffusion-guided Optimization

2.1 SDS Loss + Image Optimization (20 points)

Prompt	Iteration	Generated Image
a hamburger	1000
a standing corgi dog	1100
a cat with a hat	300
castle from dark souls 3	1400

2.2 Texture Map Optimization for Mesh (15 points)

Prompt	Generated Image
a cow with tiger skin
a cow with zebra skin

2.3 NeRF Optimization (15 points)

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

2.4.1 View-dependent text embedding (10 points)

Comparision to 2.3:
Compared to the results in 2.3, the geometry generated in 2.4 shows a significant improvement when trained with view-dependent text embeddings. For instance, the dog now correctly has two ears instead of three, and in the deer example, the model produces a single head instead of the two seen in 2.3.