Assignment 4

author: Yu Jin Goh (yujing)

Q1.1 3D Gaussian Rasterization

Rendered GIF:

Q1.2 Training 3D Gaussian Representations

Learning rates parameters used:
| Learning Rate Variable | Value | | :------- | :------: | | pre_act_opacities | 0.0005 | | pre_act_scales | 0.005 | | colours | 0.005 | | means | 0.0005 |

Number of training iterations: 1000

Metric	Value
PSNR	29.446
SSIM	0.937

Training progress:

Final Render:

Q1.3.1 Rendering Using Spherical Harmonics

Previous Render (Q1.1.5):

Now with view-dependent effects (Q1.3.1):

No View Dependence	View Dependence	Explanations
		You can observe that the patterns on the chair were originally yellow and diffuse when no view dependent effects were included but became gold as the different parts of the patterns on the chair emitted different colors based on the viewing angle once view dependent effects were accounted for
		In this view you can also observe that the decorations on the chair have highlights once view dependent effects are included and shadows on parts of the cloth that vary with viewing angle

Q2.1 SDS Loss + Image Optimization

Prompt	No Guidance	With Guidance
a hamburger
a standing corgi dog
a unicorn
a flying fire breathing dragon

Q2.2 Texture Map Optimization for Mesh

Prompt	Generated Texture
a white dairy cow with pink nose and black spots
a fearsome orange and black striped tiger

2.3 NeRF Optimization

Prompt	RGB	Depth
a standing corgi dog
a white unicorn
a tiger

2.4.1 View-dependent text embedding

Prompt	RGB	Depth
a standing corgi dog
a white unicorn

Its clear that the view-dependent text conditioning helps to improve the generated volume. As we can see in the generated radiance fields, the corgi now have only 2 ears instead of 3 and the unicorn has one horn instead of two. This is becuase the images are now scored against a more likely observation of the prompt from the corresponding view.