Assignment 4

1.1

Q1.1

1.2

Learning Rates: - Means: 0.001 - Scales: 0.01 - Opacities: 0.01 - Colors: 0.01 Number of Iterations: 1000 PSNR: 27.700 SSIM: 0.923
Training Progress
Final Renders

1.3

Without SH
With SH
Comparison 1
The rendering with full spherical harmonics (right) shows richer color variation and better captures view-dependent lighting effects compared to the lefä»–one, which appears more uniform color.
Comparison 2
Same here, the spherical harmonic version exhibits better rendering and more realistic light reflections on surfaces.
Comparison 3
This one is not that obvious, but still, the right one is more richer and more detailed.

2.1

"a hamburger"

Without guidance (400 iterations)

Hamburger without guidance

With guidance (1000 iterations)

Hamburger with guidance

"a standing corgi dog"

Without guidance (400 iterations)

Corgi without guidance

With guidance (1000 iterations)

Corgi with guidance

Prompt 3

Without guidance (400 iterations)

Prompt 3 without guidance

With guidance (1000 iterations)

Prompt 3 with guidance

Prompt 4

Without guidance (400 iterations)

Prompt 4 without guidance

With guidance (1000 iterations)

Prompt 4 with guidance

2.2

Textured Mesh 1
Textured Mesh 2

2.3

"a standing corgi dog"

Corgi RGB
Corgi Depth

"a cake"

Cake RGB
Cake Depth

"a dog"

Dog RGB
Dog Depth

2.4

"a standing corgi dog"

Corgi RGB
Corgi RGB VD
Corgi Depth
Corgi Depth VD

"a cake"

Cake RGB
Cake RGB VD
The view dependent version is more consistent though seems to lack some detail with same num of iterations. probbaly due to less fast convergence due to higher reguralization. But still, its more corehent view from different view points and more clean as shown in the depthmap.

Ziwen Yuan