Assignment 4: 3D Gaussian Splatting and Diffusion Guided Optimization

1.1 3D Gaussian Rasterization (35 points)

1.2 Training 3D Gaussian Representations (15 points)

1.3.1 Rendering Using Spherical Harmonics (10 Points)

GIF Comparision

Side by Side RGB Image Comparisons

1. Frame 0

Observations: For Frame 0, comparing the two approaches, we can see that the view-dependent one shows more natural lighting on the chair’s seat. The colors blend smoothly, and the shadows look softer and more realistic. In contrast, the view-independent approach shows a sharp boundary between shadow and bright areas, which doesn't seem realistic.

2. Frame 17

Observations: For Frame 17, comparing the two approaches, we can see that the view-dependent one shows deeper shading and subtle reflections along the chair’s edges, which makes the scene looks more natural. In contrast, the view-independent method appears to have fewer changes in brightness.

2.1 SDS Loss + Image Optimization (20 points)

1. Prompt: "a hamburger"

2. Prompt: "a standing corgi dog"

3. Prompt: "a chilling cat"

4. Prompt: "a croissant"

2.2 Texture Map Optimization for Mesh (15 points)

1. Prompt: "Tiger"

2. Prompt: "Galaxy"

2.3 NeRF Optimization (15 points)

1. Prompt: "a standing corgi dog"

2. Prompt: "a hamburger"

3. Prompt: "a standing cat"

2.4.1 View-dependent text embedding (10 points)

2.4.2 Other 3D representation (10 points)

2.4.3 Variation of implementation of SDS loss (10 points)